141
Chapter 7
The interpretation of single channel recordings
DAVID COLQUHOUN and ALAN G. HAWKES
1. Introduction
The information in a single channel record is contained in the amplitudes of the
openings, the durations of the open and shut periods, and in the order in which the
openings and shuttings occur. The record contains information about rates as well as
about equilibria, even when the measurements are made when the system as a whole
is at equilibrium. This is possible because, as with noise analysis, when the recording
is made from a small number of channels the system will rarely be exactly at
equilibrium at any given moment. If, for example, 50 percent of channels are open at
equilibrium this means only that over a long record an individual channel will be
open for 50 percent of the time; an individual channel will never be ‘50 percent open’.
In this chapter we shall consider how various aspects of single channel behaviour
can be predicted, on the basis of a postulated kinetic mechanism for the channel. This
will allow the mechanism to be tested, by comparing the predicted behaviour with
that observed in experiments. We shall also discuss the effect that the imperfect
resolution of experimental records will have on such inferences.
2. Macroscopic measurements and single channels
Measurements from large numbers of molecules will be referred to here as
macroscopic measurements. For example the decay of a miniature endplate current
(MEPC) at the neuromuscular junction provides a familiar example of a measurement
of the rate of a macroscopic process. Several thousand channels are involved, a large
enough number to produce a smooth curve in which the contribution of individual
channels is not very noticeable. In this case the time course of the current is usually a
simple exponential. Other forms of macroscopic measurements of rate include (a)
observations of the reequilibration of the current following a sudden change in
membrane potential (voltage jump relaxations), and (b) reequilibration of the current
following a sudden change in ligand concentration (concentration jump relaxations).
D. COLQUHOUN, Department of Pharmacology, University College London, Gower
Street, London WC1E 6BT, UK.
A. G. HAWKES, Statistics and OR Group, European Business Management School,
University of Wales, Swansea SA2 8PP, Wales.
142
D. COLQUHOUN AND A. G. HAWKES
In these cases too, it is common to observe that the time course of the current can be
fitted by an exponential curve, or by the sum of several exponential curves with
different time constants. This is exactly the result expected on the basis of the law of
mass action if (a) the system exists in several discrete states (see Colquhoun &
Ogden, 1986, for a discussion), and (b) the transitions between these states occur at a
constant rate. Before going further the terms used to define rates will be described.
Rate constants and transition rates
The transition rate between two states always has the dimensions of a rate or
frequency, viz s−1. For a unimolecular reaction that involves only one reactant (e.g. a
conformation change) the transition rate is simply the reaction rate constant defined
by the law of mass action. The law of mass action states that the rate of a chemical
reaction (with dimensions s−1) is directly proportional to the product of the reactant
concentrations at any given moment, the rate constant being the constant of
proportionality.
For a binding reaction, which is bimolecular (both receptor and free ligand are
reactants), the reaction rate constant has dimensions M−1s−1; the actual transition rate
in this case is the product of the rate constant (M−1s−1) and the free ligand
concentration (M) and will have dimensions s−1 as required. The assumption that the
transition rates are constant, i.e. they do not change with time, involves the
assumption that the free concentration does not change with time; this may not
always be true.
When the assumptions are satisfied the number of exponential components will, in
principle, be one less than the number of states (e.g. Colquhoun & Hawkes, 1977).
We must now ask how these relatively familiar measurements of rates are related to
what is observed with a single ion channel.
Consider the simplest possible example, a channel that can exist in only two states,
open and shut. The states will be numbered 1 and 2 respectively, so the rate constant
for transition from shut to open is denoted k21, and that for transition from open to
shut is denoted k12. The mechanism is conventionally written as
(2.1)
This is simple for two reasons. Firstly it is simple because there are only two states,
and secondly it is simple because these two states are distinguishable by looking at
the experimental record. We can see from the record which state the system is in at
any moment. In the simple mechanism in (2.1), the opening rate is k21 times the
fraction of shut channels (shut channels are the only reactant that participates in the
opening reaction). For a simple binding reaction, A+R [ AR (A=ligand molecule,
R=receptor), the association rate constant, k+1 , has dimensions M−1s−1 so when it is
multiplied by the free concentration of A (to which the association rate is
The interpretation of single channel recordings
143
proportional) we get the actual transition rate for the forward reaction with
dimensions s−1.
Macroscopic behaviour
If the reaction in (2.1) is perturbed (e.g. by a sudden change in membrane potential or
ligand concentration), the new equilibrium condition will be approached with a
simple exponential time course (there is one exponential because there are two
states). The rate of this process can be expressed as an observed time constant, τ, or as
its reciprocal, the observed rate constant λ=1/τ. These values are given by
τ = 1/(k12 + k21)
or
λ = k12 + k21
(2.2)
The equation that describes the time course contains an exponential term of the form
e−λt(=e−t/τ). Fig. 1 shows an example, an endplate current for which τ=7.1 ms. The
observed current at time t will be directly proportional to the fraction of channels that
are open (state 1) at time t, and this will be denoted p1(t). It is generally more
convenient to work with the fraction of channels (denoted p) than with the
concentration of channels (they differ only by a constant, namely the total
concentration of channels in the system). The fraction of channels in a given state is
also referred to as the occupancy of that state.
Observed rate constants
Notice that the observed or macroscopic time constant (τ), or its reciprocal the
Fig. 1. Endplate current evoked by nerve stimulation (−130 mV, inward current shown
downward). The observed current is shown by the filled circles; the continuous line is a fitted
single exponential curve with a time constant of 7.1 ms. Reproduced with permission from
Colquhoun & Sheridan (1981).
144
D. COLQUHOUN AND A. G. HAWKES
observed rate constant (λ) should be clearly distinguished from the transition rates
(k12 and k21 in this case), and the reaction (mass action) rate constants such as k+1, in
the underlying reaction mechanism; this distinction is even more important in more
complex cases. Not surprisingly the value of τ depends on both of the reaction
transition rates; equilibration will be rapid (τ will be small) if either the opening rate
(k21) is fast or the shutting rate (k12) is fast. Notice that this same expression holds for
τ whether the re-equilibration process involves a net increase or a net decrease in the
fraction of channels that are open.
The observed time constant can be written in terms of the individual rate constants
if we note that the fraction of open channels at equilibrium (after long, effectively
infinite time, and therefore denoted p1(∞)), is
p1(∞) = k21/(k12 + k21) ,
(2.3)
so the fraction of shut channels is p2(∞)=1 − p1(∞). Hence we can write
−1
τ = p1(∞)k−1
21 = p2(∞)k12
(2.4)
Therefore if only a small fraction of channels is open, so p2(∞) ⯝1, then, to a good
approximation, the observed value of τ gives us directly the value of one of the
underlying transition rates, the shutting rate k12. In general, however, the observed τ
tells us only about some more or less complicated combination of the transition rates
of the mechanism, not about any one of them alone.
So far the argument has shed little light on what is going on at the level of a single
channel. The consideration of single channel behaviour will provide a much more
concrete mental picture of what underlies the results given so far.
Randomness of single channel behaviour
There are several ways of interpreting the transition rates in terms of individual
molecules. These are described in more detail by Colquhoun & Hawkes (1983) in a
fairly informal way, and by Horn (1984), Colquhoun & Hawkes (1977, 1982, 1987)
in a more formal way. The regular behaviour observed with large aggregates of
molecules disappears when we look at a single molecule. The individual molecules
behave in an entirely random way. Or, to be more precise, the length of time for
which the molecule stays in one of its (supposedly) discrete states is quite random;
the properties of the states themselves may be very constant (for example the
amplitude of the current while the channel is open is usually very similar from one
opening to another of the same channel, and from one channel to another).
The randomness of lifetimes is immediately obvious when looking at single
channel records, and it should come as no surprise. Consider, for example, what
would happen if the lifetime of individual openings was a fixed constant. During a
MEPC a few thousand channels are opened almost simultaneously by a quantum of
transmitter (which then quickly disappears); these channels would all stay open for
their allotted time and then shut almost simultaneously so thc MEPC would be
rectangular in shape. This is not what happens. The nature of the variability will be
The interpretation of single channel recordings
145
discussed below. First two simple interpretations of the transition rates in (2.1) will be
given.
Rate constants as transition frequencies
The transition rates have the dimensions of frequency, and they can be interpreted as
such. For example in scheme (2.1) k12 is the frequency of open → shut transitions.
Such transitions can, of course, occur only when the channel is open so, to be more
precise, k12 is the number of shutting transitions per unit open time. The channel is
open for a fraction p1 of the time (this is the single channel equivalent of the
occupancy defined above), so the overall shutting frequency, the number of shutting
transitions per unit time for a single channel, fs say, is
fs = p1k12
(2.5)
Similarly the frequency of opening transitions for one channel, fo say, is
fo = p2k21
(2.6)
At equilibrium fo and fs will be equal, thus implying the result given in (2.3). It is
worth noticing that the rate constants do not tell us anything about how long the
transition from ‘shut’ to ‘open’ takes once it has started, only about how often such
transitions occur (see, for example, Colquhoun & Ogden, 1987). The transition itself
is supposed to be instantaneous; in fact the open-shut transition is unresolvably fast
(less than 10 µs, Hamill et al., 1981).
Non-equilibrium conditions. When the macroscopic current is changing with time,
as it will, for example after a voltage jump or concentration jump, then the opening
and shutting frequencies, fo and fs, will change with time and will not be equal. The
net excess of openings over shuttings per unit time, at time t, is
fo − fs = p2(t)k21 − p1(t)k12 = k21 − p1(t)(k12 + k21) .
This is exactly the expression that would be written down for the rate of change of the
fraction of open channels, dp1(t)/dt, by application of the conventional macroscopic
law of mass action, but now it has been derived by considering the flipping rate of
individual channel molecules. The proportionality constant that multiplies p1(t) in
this expression, (k12+k21) is precisely the macroscopic observed rate constant, 1/τ, as
in (2.2). Notice that the unequal frequencies result entirely from the fact that the
occupancies, p1 and p2, are changing with time in (2.5) and (2.6); the fundamental
transition rate constants (and hence the observed macroscopic rate constant that
depends only on them) are not varying with time, i.e. the membrane potential (or
agonist concentration) is held constant at all times following the application the step
change (or so, at least, it is assumed in the conventional analyses).
Rate constants and mean lifetime. Another concrete way to think of the meaning of
transition rates is to consider their relationship to how long the system stays in a
particular state. This length of individual open time is variable, but the mean open
time is a constant value and it can be shown (see below) that it is simply l/k12 for
146
D. COLQUHOUN AND A. G. HAWKES
scheme (2.1). Let us denote the mean lifetime by m; thus the mean open lifetime and
the mean shut lifetime (with dimensions of seconds) in (2.1) are, respectively,
m1 = 1/k12
m2 = 1/k21
(2.7)
More generally the mean lifetime in any individual state can be found by adding up
all the values for transition rates that lead out of that state and then taking the
reciprocal of this sum.
By combining the results in (2.7) and (2.3) we find that the equilibrium fraction of
open channels can be written in the form
m1
p1(∞) = –––––––
m1 + m2
(2.8)
i.e. it is simply the fraction of the time for which a channel is open, as asserted earlier.
Also, from (2.4), the observed macroscopic time constant can be related to the mean
lifetimes of the two states as follows:
τ = p1(∞)m2 = p2(∞)m1 = 1/(m1−1 + m2−1) .
(2.9)
Thus, if most channels are shut (p2 ≈ 1) then τ is approximately the mean open
lifetime, and if most channels are open (p1 ≈ 1) τ is approximately the mean shut
lifetime.
Rate constants and probabilities. Clearly the transition rate for shutting, k12, tells
us something about the probability that an open channel will shut. But k12 has
dimensions of s−1 so it cannot itself be a probability (which must be dimensionless).
In fact the required relationship (approximately)
Prob[open channel will shut in a small interval ∆t]=k12 ∆t
(2.10)
as long as ∆t is small (see, for example, Colquhoun & Hawkes, 1983, for more
details).
Numerical example. Suppose that in scheme (2.1) the opening transition rate is
k21=250 s−1 and the shutting transition rate is k12=1000 s−1. Thus, from (2.3), 20
percent of channels will be open and 80 percent shut at equilibrium (on average). The
observed macroscopic time constant for re-equilibration (from equations (2.2), (2.4)
or (2.9)) would be τ=1/1250=0.8 ms. The fraction of open channels is not low enough
for this to be very close to the mean open lifetime, 1/k12=1 ms. The mean shut time is
1/k21=4 ms. From (2.5) and (2.6), each channel would open (and shut) 200 times per
second (on average) at equilibrium.
3. The distribution of lifetimes
At this stage we must consider more carefully the nature of the variability of
lifetimes. Because lifetimes are random variables their behaviour must be described
The interpretation of single channel recordings
147
by probability distributions, and it is this necessity that makes rather unfamiliar the
form in which the information is contained in a single channel record. First a
distribution will be derived, and then used to illustrate a discussion of the nature of
distributions and density functions.
The distribution of random lifetimes
The first thing to define is what we mean by ‘random’. In the present context what we
mean is that the probability that an open channel will shut during a short time interval
(∆t) is constant, viz. k12∆t from equation (2.10) regardless of what has gone before
(e.g. regardless of how long the channel has already been open). And events that
occur in non-overlapping time intervals are independent of each other. In other words
what happens in the future depends only on the present state of the system, and not on
its past history. Statisticians call processes with this characteristic of ‘lack of
memory’ homogeneous Markov processes (named after the Russian mathematician
A. A. Markov, 1856-1922, who first studied them).
The conventional derivation of the required distribution starts with equation (2.10);
it is given in full by Colquhoun & Hawkes (1983) so it will not be repeated here.
Instead an alternative derivation will be given which, if less rigorous, provides
greater physical insight (an even simpler version of the following argument is given
by Colquhoun & Hawkes, 1983, p.142). This derivation can be regarded as an
exploitation of the Markov characteristics of a series of tosses of a coin. The
probability of getting ‘heads’ at each toss is the same, regardless of what has
happened before (e.g. regardless of how many consecutive ‘heads’ have been
observed in earlier tosses). The relevant analogy to tossing the coin is the random
thermal movements of the channel protein molecule. It is the randomness of thermal
motions that underlies the randomness of the lifetimes. The bonds of the protein will
be vibrating, bending and stretching, and much of this motion will be very rapid - on a
picosecond time scale. One can imagine that each time the open channel molecule
‘stretches’ there is a chance that all its atoms will get into a position where the
molecule has a chance to surmount the energy barrier and flip into the shut
conformation. Define p as the probability that the one channel will shut during each
‘stretch’, so (1−p) is the probability that it will fail to shut. Suppose that this
probability is the same at each attempt (each ‘stretch’) regardless of what has gone
before. How many attempts (stretches) will be needed before the channel eventually
shuts? The problem is just like asking how many tosses will be needed before the first
‘heads’ is encountered. The probability of success on the first attempt is p, and the
probability of failure on the first attempt but success on the second is the product of
the separate probabilities (1−p)p. The probabilities can simply be multiplied because
the two attempts are supposed to be independent of each other. The probability P(r),
of success (i.e. the channel closing) at the rth attempt (i.e. r−1 failures followed by a
success) is therefore
P(r) = (1 − p)r−1p,
r = 1, 2, . . . ∞
(3.1)
148
D. COLQUHOUN AND A. G. HAWKES
This is called a geometric distribution and its mean, µ say, the mean number of
attempts before the channel closes, is
µ = ΣrP(r) = 1/p
(3.2)
Now the ‘stretching’ is on a picosecond time scale, whereas the channel stays open
for milliseconds, so p is clearly small, and many ‘attempts’ will be needed before the
channel shuts (c.f. coin tossing, for which p=0.5). Suppose further that a ‘stretch’
occurs every ∆t picoseconds (this itself will be random but we shall treat it as though
it were constant for the purpose of this argument - doing so does not give a wrong
result). The time taken for r attempts is therefore t=r∆t, and the mean length of time
before closure is µ∆t=∆t/p. Furthermore, from the discussion above we know that the
probability of closure during ∆t is just p=k12∆t, so the mean length of time before
closure, ∆t/p, is simply 1/k12, the mean open channel lifetime found before. When r is
large we can put t=r∆t and write (3.1) in the form
Prob[channel closes between t and t + ∆t] = P(r)
≈ k12∆t(1 − p)r
= k12∆t(1 − p)k12t/p
(3.3)
Now when ∆t is small this geometric expression approaches an exponential form. The
‘compound interest formula’ states that (1+1/n)nx approaches ex as n becomes very
large (see, for example, Thompson, 1965), where e is the universal constant, the base
of natural logarithms, 2.71828.... This formula can be rearranged to show that (1−
p)x/p approaches e−x as p becomes very small. Application of this form to (3.3) gives
Prob[lifetime between t and t + ∆t] = k12∆te−k12t
(3.4)
as long as ∆t is small.
In order to express this result as a probability density function (p.d.f.) we must
divide by ∆t to give the exponential probability density function
f(t) = k12e−k12t
(3.5)
This curve is shown in Fig. 2C. Its mean is 1/k12 (see, for example, Colquhoun, 1971,
appendix 1), which proves the result for the mean open lifetime already given in
equation (2.7). The area under this curve between 0 and t gives the probability that a
lifetime is less than the specified t. This is the cumulative exponential distribution,
denoted F(t) and is found by integration of (3.5) to be
F(t) = 1 − e−k12t
(3.6)
The last two steps, and the meaning of distributions and p.d.f.s in general will next be
explained in a little more detail, before describing further the characteristics of the
exponential distribution.
Distributions and density functions
Clearly it makes sense to talk about the probability that a lifetime is less than some
The interpretation of single channel recordings
149
specified length, t; this is just the long-run fraction of lifetimes that are less than t. It is
referred to as the cumulative distribution, or distribution function, and is usually
denoted F(t). It starts at zero for t=0 and rises towards 1 for long enough times. Thus
Prob[lifetime < t] ≡ F(t) .
(3.7)
For the exponential distribution this is 1−e−k12t as shown in (3.6). Clearly it also
makes sense to talk about the probability that a lifetime lies in the interval ∆t between
two specified time values, t and t+∆t. Again this is the fraction of lifetimes that lie in
this range in the long run, and it is given, in the above notation by
Prob[lifetime between t and t + ∆t] = F(t + ∆t) − F(t) .
(3.8)
Our experimental estimate of this is what goes into the histogram bin that stretches
from t to t+∆t (see Chapter 6).
What does not make sense is to ask what is the probability of the lifetime being t,
e.g. the probability that it is exactly 5 ms, as opposed to 4.999 ms, or 5.001 ms say (we
are dealing with a theoretical distribution so the question of how precisely values can
be estimated experimentally does not arise). When ∆t is made small then (3.4) and
(3.8) approach zero; there is virtually no chance of seeing an exactly specified value.
It is for this reason that we have to introduce the idea of a probability density function
(p.d.f.). When ∆t approaches zero, so t and t+∆t converge on a single value, then (3.4)
and (3.8) also approach zero. However the ratio of these two quantities does not. The
p.d.f. is defined as this ratio, and is usually denoted f(t). This ratio is
dF(t)
F(t + ∆t) − F(t)
f(t) = –––––––––––––– → –––––
dt
∆t
(3.9)
For the exponential distribution this gives the result presented in equation (3.5)
above. When ∆t is very small this function becomes dF(t)/dt, the first derivative of the
distribution function. The p.d.f. is clearly not itself a probability (its dimensions are
s−1), but it is a function such that the area under the curve (which is dimensionless)
between any specified time values is the probability of observing a lifetime between
these values.
The form of the variability is completely specified by either the cumulative
distribution or the p.d.f., but the latter should normally be used (see comments in
chapter on fitting of data). We now return to explore the properties of the exponential
distribution in more detail.
Some properties of the exponential distribution
The exponential p.d.f. given in (3.5) has a rather unusual shape compared with other
p.d.f.s that may be more familiar. For example Fig. 2A shows the symmetrical bellshaped p.d.f. of the Gaussian (‘normal’) distribution.
Symmetry ensures that the mode (peak, or ‘most frequent value’), the median
(below which 50 percent of values lie) and the mean are all the same for the Gaussian.
Fig. 2B shows a lognormal p.d.f. (for which the log of the variable is Gaussian); the
150
D. COLQUHOUN AND A. G. HAWKES
Fig. 2. (A) The p.d.f. for a Gaussian (or ‘normal’) distribution. This example has a mean of 1.0
and a standard deviation of 0.32. (B) The p.d.f. for a lognormal distribution (a distribution
such that log x is Gaussian). (C) The p.d.f. for the exponential distribution (eqn (3.5)). The
abscissa shows the value of the variable (e.g. open lifetime) expressed as a multiple of its
mean.
The interpretation of single channel recordings
151
positive skew of this p.d.f. means that mean > median > mode. Fig. 2C shows the
exponential p.d.f. of lifetimes which is seen to be an extreme case of a positively
skewed distribution. The modal lifetime is zero. The mean of the exponential
distribution is seen to be what we would call the time constant if the exponential
curve in Fig. 2C were describing a decaying current rather than a p.d.f. If we call the
mean lifetime τ, then we can write the exponential p.d.f., (3.5), in an alternative form
as
f(t) = τ−1e−t/τ
(3.10)
The median of this distribution is 0.693τ (see below). Notice that in our simple
example τ=1/k12 is not the same as the time constant, 1/(k12+k21), from macroscopic
measurements (they will be similar only if most channels are shut, as shown above).
The single channel measurement is simpler in that it gives directly one of the
underlying reaction transition rates (k12) whereas the macroscopic measurement
gives only a combination of them.
An exponential current decay and an exponential p.d.f. are quite different sorts of
things, but it is not a coincidence that they both have the same shape. Fig. 3 shows a
simulation of what happens during the decay of an MEPC. Many channels (five of
which are shown in Fig. 3) are opened almost simultaneously by the transmitter, and
each of them stays open for an exponentially-distributed length of time before
closing.
Once closed they will not re-open because the transmitter concentration falls
rapidly to zero - this means that opening rate is zero, so in this particular case, though
not in general, the macroscopic time constant corresponds to the mean open lifetime.
At any particular moment, the total current that flows is proportional to the total
number of open channels, and this is shown in the lower part of Fig. 3. It shows an
exponential decay of the current. This is expected because the fraction of channels
that are open at t will be those that have a lifetime longer than t, i.e. 1−F(t)=e−t/τ (see
equation (3.6)). The relationship between the smooth macroscopic behaviour and the
underlying random behaviour of single channels should now be clear.
The median lifetime for an exponential p.d.f., 0.693τ, is defined such that 50
percent of lifetimes are shorter than the median and 50 percent are longer. This is the
single-channel equivalent of the half-time for the decay of the MEPC. The proportion
of lifetimes that a shorter than the mean is 63.2 percent, so 36.8 percent of intervals
are longer-than-average lifetimes. However, although longer-than- average lifetimes
are in a minority, their length is such that they actually occupy a greater proportion
(75 percent) of the time than is occupied by the more numerous shorter-than-average
lifetimes (see Colquhoun, 1971, for details). This fact means that if we stick a pin into
the record at random we will, in the long run, tend to pick out longer-than average
intervals. In fact those picked out will have, in the long run, twice the mean lifetime (a
phenomenon known as length-biased sampling). It is this rather curious property that
is responsible for some of the unexpected characteristics of random lifetimes.
The waiting time paradox. The property of length-biased sampling explains, for
example, why we get the same macroscopic time constant whether the openings of
152
D. COLQUHOUN AND A. G. HAWKES
channels are all lined up at t=0 (as in the MEPC illustrated in Fig. 3), or whether the
channel is already at equilibrium with agonist before agonist is suddenly removed at
t=0. In the latter case (but not the former) every channel that was open at t=0 (those
whose shutting is subsequently followed) will already have been open for some time
so one might think that half of their average lifetime had already been ‘used up’. In a
sense it has, but since the particular channels that were open at t=0 had a twicenormal lifetime because of length-biased sampling, the amount of time they have left
to stay open is just on average, the overall mean open time, just as in the case of the
MEPC.
The same sort of phenomenon explains why the distribution of the duration of the
shut time that precedes the first opening after a voltage jump (the first latency) will be
the same as the distribution of any other shut time. This is true, at least, for the simple
mechanism discussed above, which has only one sort of shut state (it is not, however,
always true - see discussion of correlations in section 6).
This waiting-time paradox, and other unexpected characteristics of random
Fig. 3. Simulation of the decay of an MEPC. The upper part shows five individual channels.
They have exponentially distributed lifetimes with a mean of 3.2 ms (and a median of 2.2 ms).
The lower part shows the sum of these five channels with a smooth exponential curve
superimposed on it. The curve has a time constant of 3.2 ms and a half-decay time of 2.2 ms.
Reproduced with permission from Colquhoun & Hawkes (1983).
The interpretation of single channel recordings
153
lifetimes are discussed in greater detail by Colquhoun (1971) and Colquhoun &
Hawkes (1983).
The sum of several intervals - the idea of convolution
There are many occasions when it is necessary to consider not just a single
interval, but the sum of several intervals. For example, when there are several shut
states that intercommunicate directly (as in mechanism (4.1) below), a single
observed shut period will generally consist of several oscillations between one
shut state and another - the duration of the observed shut period will be the sum of
the durations of all these (exponentially-distributed) sojourns. Similarly the
duration of a burst of openings (e.g. section 5) will be the sum of each of the open
and shut times that make up the burst. Another example occurs when the time
course of synaptic currents is considered (see section 7 below); it is then important
to consider the sum of the latency until the first channel activation plus the
duration of the activation.
The distribution of the sum of several random intervals can be found by a method
known as convolution. This method is central to much of single channel analysis, so
the technique will be introduced here by the simplest case of its use.
An example of convolution: the distribution of the sum of two intervals. We shall
derive the distribution of the sum of two exponentially distributed intervals. In the
context of the simplest mechanism (2.1), we can derive the distribution of the sum of
an open time and a shut time, i.e. the time from the start of one opening to the start of
the next. Consider an individual open time; it is an exponentially- distributed
variable, with mean 1/k12 ; its p.d.f. is
f1(t) = k12e−k12t
(3.11)
Similarly the mean shut time is 1/k21, and its p.d.f. is
f2(t) = k12e−k21t
(3.12)
Suppose that we now make a new variable by adding an open time and a shut
time. These new values will, of course, be variable; their mean value will
obviously be the sum of the separate means, (1/k12)+(1/k21), but we wish to know
what sort of distribution (p.d.f.) describes these new values. The p.d.f. of this
variable, denoted f(t) say, can be found by noting that any specified duration, t, of
the sum, can be made in all sorts of different ways. If we denote the length of the
open time as t1, and the length of the shut time as t2, then the total length is t=t1+t2.
Now if the open time happens to have a length t1=u, then, in order for the total
length to be t, the length of the shut time must be t−u. And, for any fixed value, t,
of the total length, u can have any value from 0 up to t (the former corresponds to
the case where the total length consists entirely of t2, the latter to the case where it
consists entirely of t1). The probability density that t1=u and t2=t−u can be found
by multiplying the separate probability densities (because these two events are
independent), i.e. from (3.11) and (3.12) it is f1(u)f2(t−u). Since any value of u
154
D. COLQUHOUN AND A. G. HAWKES
(from 0 to t) will do, we must now add these probabilities for each possible value
of u: since u is a continuous variable this addition takes the form of an integration,
and the result is
f(t) =
兰 u = 0 f1(u)f2(t − u) du
u=t
(3.13)
and, in this case the integral can easily be evaluated explicitly to give (as long as k12
and k21 are different)
k12k21
f(t) = –––––– (e−k12t − e−k21t) .
k21−k12
(3.14)
This is plotted in Fig. 4.
Notice that, unlike the simple exponential distribution, this distribution goes
through a maximum, at
ln(k21/k12)
tmax = ––––––––– .
k21 − k12
(3.15)
In other words the most common (modal) values are no longer the shortest values,
but values around tmax. This is not surprising because it is unlikely that t1 and t2
will both be very short. The distribution in (3.14) consists of two exponential
terms with opposite signs; one exponential represents the rising phase of the
distribution, and one the decay phase. Notice that (3.14) is completely unchanged
if k12 and k21 are interchanged: whichever is the faster represents the rising phase,
and the slower represents the decay phase. In other words the distribution of the
sum of two intervals does not depend on whether the ‘short one’ or the ‘long one’
comes first.
Equations with the form of (3.13) are known as convolution integrals. Such
equations occur whenever we consider the distribution of the sum of random
variables, and are the basis for obtaining things like the distribution of burst lengths,
though in cases more complex than this it is necessary to use Laplace transforms and
matrix methods to obtain useful results (e.g. see Colquhoun & Hawkes, 1982).
(Incidentally, convolution integrals also occur in single channel work in a rather
different context: in linear systems such as electronic filters, the output of the system
can be found by convolving the input with the impulse response function for the
system, i.e. the input is represented as a series of impulse responses, the outputs for
which superimpose linearly; see, for example, Chapter 6, and Colquhoun &
Sigworth, 1983.)
Summary of results so far
The length of time for which the channel stays in an individual state of the system (the
lifetime of this state) is a random variable. The variability is described by an
The interpretation of single channel recordings
155
Probability density (ms−1)
0.10
0.05
0.00
0
10
20
30
Time (ms)
Fig. 4. Plot of two exponential distributions with rate constants k12=1 ms−1 (mean lifetime=1
ms) (dotted curve), and k21=0.1 ms−1 (mean lifetime=10 ms) (dashed curve). The former curve
goes off-scale (intercept on the y axis is 1.0 at t=0). The solid line shows the convolution of
these two distributions, from (3.14), i.e. the distribution of the sum of a ‘1 ms’ amd a ‘10 ms’
interval.
exponential p.d.f., f(t)=τ−1e−t/τ, the mean channel lifetime, τ, being the reciprocal of
the sum of all transition rates for leaving the state in question. The exponential
distribution of lifetimes is what underlies the exponential reequilibration seen in
macroscopic measurements. The relationship between the observed time constants
and the reaction transition rates in the underlying mechanism is indirect, but will
usually (not always) be simpler for single channel measurements than for
macroscopic measurements.
So far the only concrete example that has been considered, scheme (2.1), was
simple because it had only two states, and because these two states could be
distinguished on an experimental record. We must now consider what happens when
there are more than two states, and, in particular, when not all of the states can be
distinguished from one another by looking at the experimental record.
4. Some more realistic mechanisms: rapidly-equilibrating steps
We shall consider two simple mechanisms, each of which has two shut states and one
open state.
156
D. COLQUHOUN AND A. G. HAWKES
An agonist activated channel
Consider, for example, a mechanism in which an agonist (A) binds to a receptor (R),
following which an isomerization to the active state (i.e. the open channel, R*) may
occur. This scheme, first proposed by Castillo & Katz (1957), can be written
(4.1)
The rate constants are denoted by the symbols on the arrows, and KA=k−1/k+1, is the
equilibrium constant for the initial binding step. In fact, for many fast transmitters,
more than one agonist molecule must be bound to open the channel efficiently, but
the simple case in (4.1) will suffice to illustrate the principles. It has three states, and
the lifetime in each of these states is expected to be exponentially distributed with,
according to the rule given above, the mean lifetimes for states 1, 2 and 3 being
m1 = 1/α
m2 = 1/(β + k−1)
m3 = 1/k+1xA
(4.2)
respectively (where xA is the free concentration of the agonist). The fraction of
channels that are open at equilibrium, p1(∞), is
(xA/KA)(β/α)
p1(∞) = –––––––––––––––––
1 + (xA/KA)(1 + β/α)
(4.3)
Now we can, in principle, measure the distribution of open lifetimes from the
experimental record (but see discussion of bursts, and of missed events, below); the
mean open time will provide an estimate of α directly. However a new problem arises
when we consider shut times, because the two sorts of shut state (states 2 and 3)
cannot be distinguished on the record (again see below). We therefore cannot say
which of them a shut channel is in, so we cannot measure separately the mean
lifetimes of states 2 and 3 in order to exploit directly the relationships in (4.2). This
problem will be dealt with in §5.
A channel block mechanism
Another example with two shut states and one open state (but connected differently)
The interpretation of single channel recordings
157
is provided by the case where a channel, once open, can be subsequently plugged by
an antagonist molecule in solution. This can be written
(4.4)
By the same rule as before the mean lifetimes of sojourns in open, blocked and shut
states will be, respectively,
m1 = 1/(α + k+BxB)
m2 = 1/ k−B
m3 = 1/β′,
(4.5)
where xB is the free concentration of the antagonist. Once again, the two nonconducting states (shut and blocked) cannot be distinguished by inspection of the
experimental record.
Before considering these two examples in a general way, it will be helpful to
consider what happens when several states equilibrate with each other rapidly.
Pooling states that equilibrate rapidly
The analysis becomes simpler (but also less informative) if some states interconvert
so rapidly that they behave kinetically like a single state.
Rapid binding. Consider, for example, what would happen if the binding and
dissociation of agonist in scheme (4.1) were very rapid compared with the subsequent
opening and shutting of the occupied channel. (This is probably not true for the
nicotinic acetylcholine receptor, but it makes an interesting example.) The two shut
states then appear to merge into one; this may be indicated by enclosing both in one
box, and writing the mechanism thus
(4.6)
This now looks formally just like the simple two-state case discussed first (equation
(2.1)), and can be analysed as such. (If the channel being blocked in scheme (4.4)
were an agonist activated channel then an approximation of this sort would be
implicit in (4.4), which has been written with only one shut (as opposed to blocked)
state.)
The channel shutting rate constant is α, just as in (4.1), but the opening rate
158
D. COLQUHOUN AND A. G. HAWKES
constant needs some consideration. The real rate constant for opening is β in (4.1),
the rate constant for AR→AR* transition. However, the compound ‘shut’ state in
(4.6) is sometimes occupied (AR) and sometimes not (R), but the opening
transition is possible only when it is occupied. Thus the effective opening rate
constant, β′, for transition from the compound shut state in (4.6) to the open state,
must be defined as the true opening rate constant, β, multiplied by the fraction of
time for which shut receptors are occupied (this fraction, and hence β′, will
increase with agonist concentration). Because the binding reaction has been
supposed to be rapid this fraction will essentially not vary with time (it will always
have its equilibrium value), so β′ will not vary with time, and the reaction scheme
is just like scheme (2.1). Hence, from results in (2.2) to (2.9), the open lifetimes
and the shut lifetimes will both have simple exponential distributions, with means
1/α and 1/β′ respectively. The macroscopic time constant will be τ=1/(α+β′),
which will approximate to the mean shut time if most channels are open, or to the
mean open time if most channels are shut. The equilibrium fraction of open
channels will be p1 (∞)=β′/(α+β′), which is another way of writing the result in
(4.3). There will be p1(∞)α channel shuttings (and openings) per second at
equilibrium.
How fast is fast? The answer is that it is all relative. An interesting example is
provided by a scheme like (4.4) in which the blocked state is very long- lived. An
exactly similar example would be provided by a model in which the open state
could lead to a very long-lived desensitized state (AD say). In such a case the shut
[ open equilibrium, while possibly slower than the binding reaction, might
nevertheless be much faster than the desensitization reaction. In this case all of the
states in (4.1) might be effectively pooled into a single ‘active’ (i.e. nondesensitized) state to give
(4.7)
(‘active’)
Here k−D is the rate constant for recovery from desensitization, and k′+D is the
effective rate constant for moving into the desensitized state (i.e. the true AR*→AD
rate constant, multiplied by the equilibrium fraction of active channels that are open,
and hence available to be desensitized). Again this looks just like the original two
state scheme in (2.1), and the results derived for that scheme can be applied. For
example the macroscopic rate constant for the desensitization will be τ=1/(k′+D
+k−D). The mean lifetime of the active state (each sojourn in which will consist of
many openings) will be 1/k′+D, and the mean lifetime of the desensitized state will be
1/k−D. In fact long silent desensitized periods are observed in single channel records
at high agonist concentrations (Sakmann, Patlak & Neher, 1980; Colquhoun, Ogden
& Cachelin, 1986; see also Chapter 6). It often seems to be supposed that if, say, 10
The interpretation of single channel recordings
159
second silent desensitized periods appear in the single channel record then we
should see that macroscopic desensitization (at the same agonist concentration)
develops with a similar time constant. However the argument just presented shows
that, insofar as the desensitized periods are much longer than the non-desensitized
active periods, then the macroscopic time constant should be closer to the mean
length of the periods of activity than to the mean lengths of the silent desensitized
periods.
5. Analysis of mechanisms with three states
Under the usual assumptions (see above) we would expect macroscopic observations
on any mechanism with three states to reequilibrate along a time course described by
the sum of two exponentials with different time constants (τs and τf say, for the
slower and faster values respectively). In the cases just discussed, where some states
equilibrate very rapidly with each other, one of the time constants would be so fast as
to be undetectable, so we would appear to have only two states. We must now
consider the case when this does not happen.
A channel block example
Fig. 5 shows an example where two exponentials are clearly visible: it is an endplate
current recorded in the presence of a channel blocking agent, gallamine, for which τs
=28.1 ms and τf=1.37 ms. Noise spectra also show two components and the results
can be accounted for by a simple block mechanism in (4.4), as described by
Colquhoun & Sheridan (1981).
As in the simplest case the macroscopic time constants (τs, τf) or their reciprocals,
the macroscopic rate constants (λs, λf), are indirectly related to all of the transition
Fig. 5. An evoked endplate current as in Fig. 1 but recorded in the presence of a channel
blocking agent (gallamine 5 µM). The current is shown by the closed circles the fitted double
exponential curve has time constants τf=1.37 ms, τs=28.1 ms. Reproduced with permission
from Colquhoun & Sheridan (1981).
160
D. COLQUHOUN AND A. G. HAWKES
rates in the underlying mechanism. The macroscopic rate constants in this case are
found by solving a quadratic equation,
λ2 + bλ + c = 0 .
(5.1)
The well-known solution of this quadratic is
λs, λf = 0.5共− b ± √b2 − 4c兲.
(5.2)
A less well-known alternative is
2c
λs, λf = –––––––––––– ,
− b ⫿ √b2 − 4c
(5.3)
− b = λs + λf
(5.4)
c = λ sλ f .
(5.5)
where
and
At this point it will be useful to point out that when one of the rate constants is much
bigger than the other (say λf » λs , where the subscripts denote ‘fast’ and ‘slow’), i.e.
when b2 » c, then it follows from (5.2) and (5.4) that the faster rate constant, λf, is
approximately
λf ≈ − b
(5.6)
and, from (5.3) and (5.5), the slower rate constant, λs, is approximately
λs ≈ − c/b
(5.7)
In this case the coefficients, b and c, are given by
− b = α + β′ + k+BxB + k−B ,
冤
冢
冣冥 ,
xB
β′
c = αk−B 1 + –– 1 + –––
α
KB
(5.8)
(5.9)
and KB=k−B/k+B is the equilibrium constant for binding of the blocking agent
(concentration xB) to the open channel. More generally when there are n states the n−
1 macroscopic rate constants will be solutions of a polynomial of degree n−1, and
their sum will be equal to the sum of all the underlying reaction transition rates, as in
(5.4). A third parameter, the relative amplitudes of the fast and slow components, is
needed to describe completely results such as those in Fig. 5. This too is related (in a
still more complex way) to the reaction transition rates; it can be calculated by
methods such as those described by Colquhoun & Hawkes (1977). It is a very
The interpretation of single channel recordings
161
important stage in the testing of a putative mechanism to ensure that not only the time
constants, but also their amplitudes, are as predicted by the mechanism.
In order to test the fit of the channel block mechanism, (4.4), to experimental data
we can, for example, see whether λs and λf (or, more conveniently, λs+λf and λsλf),
and the relative amplitude of the two components, vary with the concentration of
blocker (xB) and the concentration of agonist (which changes β′ - see above) in the
predicted way. If it does then values for the underlying transition rates, α, k−B etc. can
be extracted; in this process it is often useful to use approximations to the exact
equations (such as (5.6) and (5.7)), and various procedures have been described (see,
for example, Adams, 1976; Adams & Sakmann, 1978; Colquhoun, Dreyer &
Sheridan, 1979; Colquhoun & Sheridan, 1981).
A simple agonist example
Similar arguments apply to the simple agonist mechanism in (4.1). Again two
macroscopic rate constants are predicted, their values being found from the quadratic
solution (5.2), but in this case the coefficients are given (e.g. Colquhoun & Hawkes,
1977) by
− b = λs + λf = α + β + k+1xA + k−1
(5.10)
xA (α + β)
c = λsλf = αk−1 1 + –––
–––––––
α
KA
冤
冥
=αk−1 + αk+1xA + βk+1xA
(5.11)
where KA is the microscopic equilibrium constant for the binding reaction, i.e.
k−1/k+1. The problem with trying to use this result is that it is not generally possible to
see experimentally the predicted two components in macroscopic responses to
agonists. Usually, as in Fig. 1, a single exponential is a good fit. It is for this reason
that it was postulated that agonist binding is very fast, as described in (4.6). However
this is not the only possible explanation (and is probably not the correct explanation).
In order to get further, the greater discriminating power of single channel
measurements are needed.
Single channel results with more than two states
In practice it is usually observed, for virtually every type of channel, that the p.d.f.
required to fit the distribution of shut times has more than one exponential
component. For example Colquhoun & Sakmann (1985) found that three components
were needed for the nicotinic channel even at low agonist concentrations (see Fig. 10
in Chapter 6). Under the usual assumptions, the number of components in this
distribution provides a (minimum) estimate of the number of shut states (see
Colquhoun & Hawkes, 1982). Likewise the number of components in the open time
distribution indicates the (minimum) number of open states. These results are rather
more informative than for macroscopic measurements for which the number of
components (minus one) indicates the total number of states, but doesn’t indicate
how many of these are open states.
162
D. COLQUHOUN AND A. G. HAWKES
Thus both the examples above, the agonist mechanism in (4.1) and the channel
block mechanism in (4.4), predict a double exponential shut time p.d.f. and a single
exponential open time p.d.f. The open time p.d.f. will therefore be
f(t) = τ−1e−t/τ
(5.12)
where τ is the mean open time given in (4.2) and (4.5) for the two models (notice that
this is decreased by increasing the blocker concentration for the channel block case).
The shut time p.d.f. should have the double-exponential form (see Chapter 6)
f(t) = asτs−1e−t/τs + afτf−1e−t/τf ,
(5.13)
where as and af are the relative areas under the p.d.f. accounted for by the slow and
fast components respectively, and τs and τf are the observed ‘time constants’, or
‘means’, of the two components. We may also refer to their reciprocals, the observed
rate constants for the shut time distributions, λs=1/τs and λf=1/τf. These time
constants are not, of course, the same as those found for macroscopic measurements.
We must, therefore, now enquire how they are related to the underlying reaction
transition rates.
Shut times for channel block. This case is particularly simple because no direct
transition is possible between the two sorts of shut state in (4.4). Every shutting of the
channel must, therefore, consist of either a single sojourn in the shut state (state 3), or
a single sojourn in the blocked state (state 2). The two time constants in this case are
therefore simply the mean lifetimes of these two states, m3=1/β′ and m2=1/k−B (the
mean lifetime of a blockage). And the relative areas are determined simply by the
relative frequency of sojourns in each of these states, i.e. on the relative values of α
and k+BxB. These results are a great deal simpler than those for macroscopic
measurements given in (5.8)-(5.9); the values of two transition rates can be found
directly from τs and τf.
Shut times for the agonist mechanism. The scheme in (4.1) is a bit more
complicated than that for channel block; the reason for this is that the two shut states
intercommunicate directly. Thus an observed shut period might consist of a single
sojourn in AR (state 2) followed by reopening of the channel, but it might also consist
of a variable number of transitions from AR to R and back before another opening
occurred (both states are shut so these transitions would not be visible on the
experimental record). This complication means that we once again have to resort to
solving a quadratic, as in (5.2), to predict the observed time constants for the shut
time distribution. One way in which the distribution of shut times can be derived, for
this particular mechanism, is given in appendix 1. By use of the matrix methods given
by Colquhoun & Hawkes (1982), a general expression for the shut time distribution
can be obtained that holds for any mechanism. In this case the observed rate constants
are given by (e.g. appendix 1, and Colquhoun & Hawkes, 1981) by solving the
quadratic, as in (5.1)-(5.7),
−b = λs + λf = β + k+1xA + k−1
(5.14)
The interpretation of single channel recordings
c = λsλf = βk+1xA .
163
(5.15)
This result, although more complex than for the channel block mechanism, is still a
good deal simpler than the analogous result for macroscopic measurements given in
(5.10) and (5.11). For example, the rate constant α does not appear at all here (only
those transition rates that lead away from shut states appear), and the expressions are
simpler. Nevertheless, as in the macroscopic case, the observed time constants are not
directly related to individual reaction transition rates or to the mean lifetimes of
states.
Although this statement is generally true, we can often find good approximations
that lend a simple physical significance to the time constants. If, for example, we
observe that one rate is very much faster than the other (λf » λs) then (5.14) will give a
good approximation to λf; furthermore when the agonist concentration (xA) is very
low the term k+1xA will be negligible, so under these conditions
1
τf = 1/λf ≈ ––––––– ,
β + k−1
(5.16)
and this is just the mean lifetime of a single sojourn in the AR state, as was pointed
out in (4.2). Note, though, that while this is exactly the mean lifetime of AR, it is only
approximately the fast time constant of the shut time distribution. The physical
significance of this is that short shut periods (gaps) consisting of a single sojourn in
AR are predicted to appear in between openings of the channel as a result of
AR*→AR→AR* . . . transitions. If, while in AR, the agonist dissociates (AR→R),
rather than the channel reopening (AR→AR*), then a much longer shut period will
ensue, which will contribute to the slow component of the distribution of shut times.
Thus the openings will appear to occur in rapid bursts.
Bursts of channel openings
Whenever the shut time distribution has one time constant that is much shorter than
another the experimental record will contain short shut periods and some much
longer ones (see, for example, the shut time distribution shown in Fig. 10 of Chapter
6). In this case several openings may occur, each opening separated from the next by
a short shut period, before a long shut period occurs. The openings, will therefore
appear to be grouped together in bursts of openings, each burst being separated from
the next by a relatively long silent period. Examples of such bursts are shown, for
example, by Neher & Steinbach (1978), Ogden, Siegelbaum & Colquhoun (1981),
and Ogden & Colquhoun (1985) for channel blockers and by Colquhoun & Sakmann
(1981, 1983, 1985) and Sine & Steinbach (1985) for agonists. Some bursts are
illustrated in Fig. 6.
It may simplify the analysis considerably if the experimental record can be
divided up, without too much ambiguity, into such bursts. A major advantage of
doing this is that we can usually be sure that all the openings within one burst
originate from the same individual channel, so the characteristics of shut periods
164
D. COLQUHOUN AND A. G. HAWKES
Fig. 6. (A) Burst-like appearance of single channel currents in different transmitter-activated
ion channels. (Ai) Acetylcholine-activated single channel current recorded from the frog
endplate. Downward deflection of the trace represents inward current. (Aii) Acetylcholineactivated single channel current recorded from a sinoatrial node cell of rabbit heart.
Downward deflection represents inward K+ current. (Aiii) Glycine-activated single channel
current in soma membrane of a cultured spinal cord cell. Upward deflection of the trace
corresponds to an outward current reflecting Cl− influx into the neuron. Reproduced with
permission from Colquhoun & Sakmann (1983). (B) Bursts produced by rapid channel block.
(Bi) Acetylcholine activated channels at frog end-plate. (Bii) The same in the presence of
benzocaine (200 µM). Reproduced from Ogden, Siegelbaum & Colquhoun (1981).
The interpretation of single channel recordings
165
within bursts can tell us something about the channel. On the other hand individual
well-separated bursts may well originate from different channels so, in the absence
of knowledge of how many channels there are in the membrane patch, it is
impossible to interpret the lengths of gaps between bursts in any useful way. Another
advantage of looking at the properties of bursts of openings is that the interpretation
of the results may be simplified; long-lived shut states are, by definition, excluded
from the bursts so we need consider a smaller subset of states than would otherwise
be necessary.
Bursts with channel blockers
The appearance of bursts in the presence of a channel blocking agent will depend on
how long the blockages last. If the blockages last for a long time (e.g. seconds, i.e.
k−B is small) the blocked periods will not be distinguishable from any other shut time,
and the record will not look obviously bursty. At the other extreme, if blockages are
very brief (i.e. k−B is very large) so that they are mostly not resolvable, it will appear
(incorrectly) that the single channel amplitude is depressed; this happens partly
because such blockers will have a low affinity for the receptors, and so will be used in
concentrations that make the openings very short (see (4.5)), as well as the blockages
being short (see Ogden & Colquhoun, 1985, for an example).
Clear bursts will be seen in the presence of channel blockers when the blockages
are of a duration that is comparable with the channel open times, so blockages
appear as brief interruptions of a normal channel activation. Each individual opening
will be shorter in the presence of the blocking agent, as shown in (4.5), and the
predicted linear dependence of the reciprocal open time on blocker concentration
provides a test of the model (and, if the test is passed, an estimate of the association
rate constant for block of an open channel, k+B). An example of the linear
dependence of the reciprocal mean open time on blocker concentration is shown in
Fig. 7 (filled circles).
The line is straight as predicted by (4.5) so the slope provides an estimate of the
association rate constant for the blocking reaction (k+B=3.9 × 107 M−1 s−1 in Fig. 7).
In this case the mean open times had been corrected for the fact that the brief
interruptions, which the agonist (suberyldicholine) shows even in the absence of
block, are only partially resolved. The intercept should therefore give an estimate of
α. Another approach would be to ignore these spontaneous shuttings and look at the
shortening of whole elementary event (the burst of openings produced by a single
activation of the channel) by channel blocking; this should also give a slope of k+B
(Ogden & Colquhoun, 1985).
Fallacies in the interpretation of channel block
The shortening of the mean open time by a channel blocker might be regarded as
resulting from openings being ‘cut short’ by the insertion of a blocking molecule into
the open channel. This view seems consistent with the fact that the mechanism (4.4)
predicts that the mean value of total open time in a burst is unaltered by the presence
166
D. COLQUHOUN AND A. G. HAWKES
4500
4000
rate/s−1
3000
2000
1000
0
10
50
concentration of SubCh/µM
100
Fig. 7. Block of suberyldicholine-activated end-plate channels by the agonist itself at high
concentrations. The blockages produce a characteristic component with a mean of about 5 ms
in the distribution of shut times (so k−B≈1/5 ms=200 s−1). Blockages become more frequent
(the relative area of the 5 ms component increases) with concentration. Closed circles:
reciprocal of the corrected mean open time plotted against concentration (slope 3.9×107 M−1
s−1, intercept 594 s−1). Open circles: the ‘blockage frequency plot’ - the frequency of
blockages per unit of open time is plotted against concentration (slope=3.8×107 M−1 s−1 ,
intercept 4 s−1). Reproduced with permission from Ogden & Colquhoun (1985).
of the blocker (it is still 1/α as it would be in the absence of the blocker). This
prediction is also useful for testing the fit of the results to the postulated mechanism
(e.g. Neher, 1983). The basis of this prediction is outlined by Colquhoun & Hawkes
(1983); it is as though the opening ‘carried on as normal’ after a blockage. However
this way of thinking of the blocking effect is clearly incompatible with a memoryless
process; the channel cannot ‘know’ how long it has been open during earlier openings
in the burst. And it can lead to errors; for example it might be supposed that if we
separated out from the record all of those channel activations in which no blockage
happened to occur then such activations would not be ‘cut short’ so they would have a
normal mean open time of 1/α (the number of blockages per activation is random,
there may be 0, 1, 2, . . . blockages; see Fig. 8). This is not true: in fact openings thus
selected would have a shortened mean lifetime of 1/(α+k+BxB) just like the openings
that were terminated by a blockage (see Colquhoun & Hawkes, 1983 for a fuller
discussion).
The blockage frequency plot
When activations of the channel are frequent (at high agonist concentrations) the
beginning and end of a burst will not be clearly distinguishable, so the prediction that
the mean total open time per burst is unaltered by the channel blocker cannot be
The interpretation of single channel recordings
167
tested. However a similar test can be achieved by doing a ‘blockage frequency plot’
(Ogden & Colquhoun, 1985), as long as a component that is attributable to blockages
can be distinguished in the shut time distribution. By analogy with (2.5), the
frequency of channel blockages should be p1k+BxB where p1 is the fraction of open
channels. Thus, if the frequency of blockages per unit open time is plotted against the
blocker concentration the slope of line should be k+B. An experimental example of
such a plot is shown in Fig. 7 (open circles); it is seen to behave as predicted by the
simple channel block mechanism. The slope of this line gives k+B=3.8×107 M−1 s−1, a
very similar value to that estimated from the shortening of openings.
Fig. 8. Simulation of MEPC decay as in Fig. 3, but in this case a channel blocker is supposed
to be present so some of the openings (all but channels 2 and 3) are interrupted by one or more
blockages. The sum of all seven channels is shown in the lower part; a double exponential
decay curve is superimposed on it (the two separate exponential components are shown as
dashed lines).
168
D. COLQUHOUN AND A. G. HAWKES
Relationship between bursts and macroscopic currents with channel block
The relationship between bursts at the single channel level, and double
exponential relaxation at the macroscopic level is illustrated by the simulated
MEPC in Fig. 8.
As in Fig. 3, the seven channels that are illustrated open almost synchronously,
but this time the opening may be interrupted by blockages before the agonist
dissociates and the channel shuts. For example, channel 1 has two such blockages,
channel 5 has three blockages, and channels 2 and 3 have no blockages. When the
channels are summed the total current is seen to decay with a double exponential
time course (as illustrated in the experiment in Fig. 5). There is an initial rapid
decay as channels shut, or are blocked for the first time, and then a much slower
decay the rate of which reflects (approximately) the length of the whole burst of
openings. The latter follows from (5.7), and the definition of b and c in (5.8) and
(5.9). These give, if the agonist concentration is small enough (i.e. β′ is small
enough), and the mean burst length is long compared with the length of a
blockage,
τs = 1/λs ≈ −b/c
1 1 + xB/KB
≈ ––– + ––––––––
k−B
α
(5.17)
The second term on the right hand side is the mean burst length, as will now be shown
(see equation 5.23). Before doing this, it is necessary to consider the number of
blockages that occur in a burst.
The number of blockages per burst. First note that the number of openings in a
burst must always be one greater than the number of blockages, so we shall actually
work with the number of openings per burst. This will, like everything else, be
random. The distribution of the number of openings per burst, and hence an
expression for the mean number, can be derived as follows. Consider an open channel
(state 1 in scheme (4.4)). Its next transition may be either blocking (going to state 2
with rate k+BxB), or shutting (going to state 3 with rate α). The relative probability of
the former happening (regardless of how long it takes before it happens), which we
shall denote π12, is thus
k+BxB
π12 = ––––––––––
(5.18)
α + k+BxB
and the probability of the latter happening is therefore π13=1−π12. Notice also that a
blocked channel (state 2) must unblock (to state 1) eventually, so π21=1. The
probability of blocking and then reopening is therefore π12π21=π12. A burst will
contain r openings (and r−1 blockages) if an open channel blocks and unblocks r−1
times (probability π12r−1 ), and then returns after the last opening to the long- lived
shut state, state 3 (with probability π13). The probability of seeing r openings (i.e. r−1
blockages) in a burst is thus
The interpretation of single channel recordings
169
P(r) = π12r−1 π13 .
(5.19)
This form of distribution is known as a geometric distribution. It is the analogue,
for a discontinuous variable, of the exponential distribution, and has the same shape
as an exponential distribution except that it decays in steps, rather than continuously.
One or other of the possible outcomes (r=1, 2, 3, . . .,∞) must occur, so these
probabilities must add to unity, i.e., from (5.19),
∞
冱P(r) = 1 .
(5.20)
r=1
The mean number of openings per burst can be found from the general expression
for the mean, µ, of a discontinuous distribution, viz.
∞
µ=
冱rP(r) .
(5.21)
r=1
In the present example, the mean number of openings per burst is therefore
k+BxB
1
µ = –––––– = 1+ –––––– .
1 − π12
α
(5.22)
As expected this depends simply on the relative probabilities of leaving the open state
for the blocked, or the shut, states. The mean number of blockages per burst, k+BxB/α,
is directly proportional to the blocker concentration.
We can now work out the mean length of a burst. The average burst consists of µ
openings, each of mean length m1, and µ−1 blockages, each of mean length m2 (as
defined in (4.5)). The mean burst length is therefore
1 + xB/KB
µm1 + (µ − 1)m2 = –––––––– .
α
(5.23)
This should increase linearly (from 1/α) with the blocker concentration.
Bursts with agonists alone
One reason for expecting openings to occur in bursts has been discussed above, viz.
multiple openings during a single occupancy. Whatever the reasons the phenomenon
is certainly common, as illustrated in Fig. 6. A typical histogram of all shut times for
an agonist activated channel is given, and discussed, in Chapter 6 (Fig. 10).
If we assume that a mechanism like (4.1) is what underlies these observations, then
we can extract information from the results by two sorts of measurements. It has
already been mentioned (4.2) that the mean length of the short gaps within a burst will
be approximately 1/(β+k−1). This alone does not allow us to estimate the separate
values of β and of k−1 . However it is also true (see below) that the mean number of
170
D. COLQUHOUN AND A. G. HAWKES
openings per burst should be 1+β/k−1. If this is measured too, then the channel
opening rate constant β, and the agonist dissociation rate constant k−1 can both be
separately estimated (e.g. Colquhoun & Sakmann, 1985).
The number of openings per burst for agonist alone. The distribution of the number
of openings per burst can be derived in much the same way as in the case of bursts
caused by channel blockages, in (5.18) - (5.22). Consider a channel in the
intermediate state AR (state 2 in scheme (4.1)). Its next transition may be either
reopening (rate β) or agonist dissociation (rate k−1). The relative probability of the
former happening (regardless of how long it takes before it happens), which we shall
denote π21, is thus
β
π21 = –––––––
β + k−1
(5.24)
and the probability of the latter happening is therefore π23=1−π21. A burst must
contain at least one opening; for the channel to re-open r−1 times before dissociation
occurs the former event must occur r−1 times before the latter happens. The
probability of r openings in a burst (r=1, 2, 3, ..., ∞) is thus
P(r) = πr−1
21 π23
(5.25)
This is the geometric distribution of the number of openings per burst. The mean
number of openings per burst is thus, from (5.21)
µ = 1/π23 = 1 + β/k−1 ,
(5.26)
as asserted above.
The actual molecular transitions that underlie the bursting behaviour in scheme
(4.1) are illustrated explicitly in Fig. 9. Notice the ‘invisible’ oscillations between the
Fig. 9. Schematic illustrations of transitions between various states (top) and the
corresponding observed single channel currents (bottom) to illustrate the origin of bursting
behaviour for the simple agonist mechanism in (4.1). Notice that many transitions between
shut states are invisible on the experimental record.
The interpretation of single channel recordings
171
two shut times (mentioned above) that result from agonist occupancies that fail to
produce opening of the channel. A more elaborate version of this diagram is given by
Colquhoun, Ogden & Cachelin (1986) who use it to illustrate what happens when
either the binding step is very fast (see above) or, alternatively, when the open-shut
conformation change is very fast.
6. Correlations in single channel records
It was assumed above that the future evolution of a system depends only on its
present state and not on its past history. For the simple open-shut model in (2.1) this
implies, for example, that the length of an opening must be quite independent of (and
therefore uncorrelated with) the length of the preceding shut time and of the lengths
of preceding open times. The same is true of the 3-state mechanisms in (4.1) and
(4.4); furthermore there will be no correlation between the length of one burst of
openings and the next for either of these mechanisms. However more complex
mechanisms may show such correlations; they are still ‘memoryless’, so sojourns in
individual states will be independent, but correlations can arise because it is not
possible to distinguish one sort of shut state from another on the record (they all have
zero conductance), and similarly open states of equal conductance are not
distinguishable. In particular, there will be correlations if there are (a) at least two
open states (b) at least two shut states and (c) at least ‘two routes’ between the shut
states and the open states (Fredkin, Montal & Rice, 1985; Colquhoun & Hawkes,
1987; Ball & Sansom, 1988).
Correlations between open times and between shut times
The last condition for the appearance of correlations, that there should be at least ‘two
routes’ between the shut states and the open states, must now be put rather more
precisely. Correlations will be found if there is no single state, deletion of which
totally separates the open states from the shut states. The number of states which must
be deleted to achieve such a separation is the connectivity of open and shut states, so
correlations will be seen if the connectivity is greater than one. The mechanisms in
(6.1) show three mechanisms each with
(6.1)
two open states (denoted O) and three shut states (denoted C). In schemes (a) and (b)
there will be no correlations; deletion of state C3 (or of state O2) in (a) separates the
open and shut states, as does deletion of C3 in (b). In (c), on the other hand, the
172
D. COLQUHOUN AND A. G. HAWKES
connectivity is 2 (e.g. deletion of C3 and C4 will separate open and shut states) so
correlations between open times may be seen. Even in this case correlations between
successive open times will be seen only if the two open states, O1 and O2, have
different mean lifetimes. The correlations result simply from the occurrence of
several C4[O1 oscillations followed by a C4→C3 transition and then several C3[O2
oscillations, so runs of O1 and runs of O2 openings occur. The effect will clearly be
most pronounced if the C4[C3 reaction is relatively slow.
For example, most of the properties of the nicotinic receptor are predicted well by
(c): in this case O2 has a long mean lifetime compared with O1 (but it has the same
conductance), whereas C3 has a very short lifetime. Thus long open times tend to
occur in runs (so there is a positive correlation between the length of one opening and
the next), but long openings tend to occur adjacent to short shuttings, giving a
negative correlation between open time and subsequent shut time Colquhoun &
Sakmann (1985).
In the earlier work in this field, it was usual to measure correlation coefficients
from the experimental record. However it is visually more attractive, and in some
respects more informative, to present the results as graphs, as suggested by
McManus, Blatz & Magleby (1985), Blatz & Magleby (1989), and Magleby & Weiss
(1990). An example of such a plot is shown in Fig. 10.
This graph illustrates correlations found for the NMDA-type glutamate receptor
(Gibb & Colquhoun, 1992). To construct this graph, five contiguous shut time ranges
were defined (each range being centred around the time constant of a component of
the shut time distribution). Then, for each range, the average of the open times was
calculated for all openings that were adjacent to shut times in this range, and this
average open time was plotted against the mean of the shut times in the range. The
Mean adjacent open time (ms)
4
3
2
1
0
0.01
0.1
1
10
100
1000
Mean specified shut time range (ms) (log scale)
Fig. 10. Relationship between the mean durations of adjacent open and shut intervals. The
graph shows the mean ± standard error of the open times in sixteen different patches, plotted
against the average of the mean adjacent shut time ranges used in each patch. Reproduced
from Gibb & Colquhoun (1992).
The interpretation of single channel recordings
173
graph in Fig. 10 shows a continuous decline, so it is clear that long open times tend to
be adjacent to short shut times, and vice versa.
In principle the connectivity between open and shut states can be measured
experimentally because the correlation between an open time and the nth subsequent
open time (for a single channel) will decay with increasing lag (n) towards zero as the
sum of m geometric terms, where m is the connectivity minus one (Fredkin et al.,
1985; Colquhoun & Hawkes, 1987). However such a quantitative interpretation of
correlations has not yet been achieved in practice.
These results can be extended to correlations between the lengths of bursts of
openings, and between the lengths of openings within a burst (Colquhoun & Hawkes,
1987). There will be correlations between bursts when the connectivity (as defined
above) between open states and long-lived shut states is greater than one. There will
be correlations between openings within a burst when the direct connectivity between
open states and short-lived shut states is greater than one (the term direct connectivity
refers only to routes that connect open and short-lived shut states directly, not
including routes that connect them indirectly via a long lived shut state, entry into
which would signal the end of a burst). Thus, for the examples in (6.1), taking C5 to
be the long lived shut state, neither (a) nor (b) would show any such correlations,
whereas (c) would show correlations within bursts but no correlations between bursts
(as observed experimentally by Colquhoun & Sakmann, 1985). The following
scheme (in which C5 and C6 both represent long lived shut states), on the other hand,
would show all three types of correlation
(6.2)
7. Single channels after a voltage- or concentration-jump
The lifetime of a sojourn in a single state will always (under our assumptions) follow
an exponential distribution. However some sorts of measurements may be dependent
on whether the recording is made at equilibrium or not. So far we have supposed that
recordings are made at equilibrium. Strictly speaking the assumption is a bit less
strong - we assume that there is a steady state; however, for the sort of mechanism we
are discussing this amounts to the same thing (Colquhoun & Hawkes, 1982, 1983).
Some new, and more complex, considerations arise when the record is not at
equilibrium. For example, following a sudden change in concentration (a
concentration-jump) or membrane potential (a voltage-jump), it will take some time
for a new equilibrium to be established. It is usual to look at the channel openings (or
shuttings, or bursts) that follow such a perturbation as though the 1st, 2nd, . . .
174
D. COLQUHOUN AND A. G. HAWKES
opening (or shutting, or burst) were all directly comparable. This will be the case if
there is only one sort of shut state and one sort of open state as in the simplest scheme
(2.1) that was discussed earlier (Colquhoun & Hawkes, 1987).
Consider, for example, the case where the channels are shut initially but may open
following a depolarizing voltage step applied at t=0. If there is more than one sort of
shut state the latency until the first opening may depend on how the system was
distributed among the various shut states at t=0, so measurements of the first latency
can give information about this, and about the lifetimes of these shut states.
Thereafter, however, every open time and shut time will be exactly comparable, each
having exactly the same distribution as it would have in an equilibrium record. This
simple result will happen in precisely those cases in which open and shut times are
not correlated (as described in the preceding section).
When open and shut times are correlated then the lengths of the 1st, 2nd, . . . etc.
openings (or shuttings, or bursts) will not all have the same distribution, though they
will approach the equilibrium distribution eventually. Their distributions should,
however, all have the same observed time constants; these time constants are
functions (though possibly quite complicated functions) only of the underlying
transition rates and, according to our assumptions, these remain constant throughout
the post-jump period, because the membrane potential and/or ligand concentrations
are supposed to remain constant. What is happening is that the occupancies of each
state, and hence, from (2.5) and (2.6), the actual transition frequencies, are varying
with time. This means that the relative areas of the components associated with each
time constant may be different for the 1st, 2nd, . . . etc. open time following the jump.
This phenomenon is discussed and exemplified by Colquhoun & Hawkes (1987), but
it has not yet been exploited experimentally.
It is worth noting that voltage jumps and concentration jumps are usually quite
brief (often in the range 1-300 ms), so they are not a good way of investigating slow
kinetic processes (such as the so-called mode changes), which are necessarilly
obscured in short recordings.
The time course of synaptic currents
In many (though possibly not all) cases it seems that the time course of the postsynaptic current elicited by nerve stimulation is quite long compared with the length
of time for which the transmitter is present in the synaptic cleft (e.g. Anderson &
Stevens, 1973). In other words the presynaptic ending provides a very brief
concentration jump of transmitter; this opens some channels which subsequently shut
in the absence of transmitter. In the simulations in Figs 3 and 8 it was supposed that
channels open essentially synchronously, i.e. that the latency to the first opening (first
latency, for short) following exposure to the agonist is very short. This is probably
close to the truth for fast channels such as the muscle-type nicotinic acetylcholine
receptor (e.g. Franke et al., 1991; Liu & Dilger, 1991), and may well be true for at
least some AMPA-type glutamate receptors in the CNS (Colquhoun, Jonas &
Sakmann, 1992).
However it is clearly not true that channels open synchronously for the NMDA-
The interpretation of single channel recordings
175
type glutamate receptor (Edmonds & Colquhoun, 1992); in this case the first opening
may occur hundreds of milliseconds after a brief pulse of glutamate is applied. The
consequences of this for the time course of the synaptic curent can be illustrated by
the following oversimplified example.
Consider a hypothetical channel which, after brief agonist application, produces an
activation consisting of a single opening, after the first latency has elapsed (for the
NMDA receptor, the activation is actually a great deal more complicated than a single
opening). In Figure 11A, nine examples are shown of simulated channels with a mean
first latency of 1 ms, and a mean open time of 10 ms (the variability of both being
described by simple exponential distributions).
The average current (shown at the top), is seen, not surprisingly, to have a rising
phase that can be fitted with an exponential with a time constant of about 1 ms, and
the decay phase has a time constant of about 10 ms. Apart from being about 10 times
too slow, this example is similar to what happens at a neuromuscular junction.
0.5 pA
Mean latency 1 ms
Mean open time 10 ms
1000 channels averaged
10 ms
B
Mean latency 10 ms
Mean open time 1 ms
1000 channels averaged
0.05 pA
A
10 ms
τ (ms) ampl (pA)
1.03
1.11
10.6
−1.11
Y(∞)=
0.0
τ (ms) ampl (pA)
0.781
0.0922
12.0 −0.0913
Y(∞)=
0.0
Fig. 11. A simple simulation of a synaptic current, in which each channel is supposed to
produce a single opening only, after an exponentially-distributed latency. (A) Mean latency 1
ms, mean open time 10 ms. The lower part shows nine examples of simulated channels. The
top trace is the average of 1000 such channels; the double-exponential curve fitted to the
average has τ=1.03 ms (amplitude 1.11 pA), and τ=10.6 ms (amplitude=−1.11 pA). (B)
Similar but with a mean latency of 10 ms and a mean open time of 1 ms. The doubleexponential curve fitted to the average has τ=0.76 ms (amplitude 0.092 pA), and τ=12.0 ms
(amplitude=−0.091 pA).
176
D. COLQUHOUN AND A. G. HAWKES
More surprising, perhaps, are the results shown in Figure 11B, in which the
numbers are reversed, and simulated channels have a mean first latency of 10 ms, and
a mean open time of 1 ms. The averaged current shown at the top is seen to have
essentially the same shape as in Figure 11A (though it is ten times smaller, and
considerably noisier relative to its amplitude). Thus, in this case, the rate of decay
reflects the duration of the first latency, whereas the rate of rise represents the mean
channel open time (this case resembles observations on sodium channels: Aldrich,
Corey & Stevens, 1983). The reason for this result, which seems paradoxical at first
sight, can be seen from the simulations (e.g. the exponential distribution of first
latencies means that short latencies are more common than long ones), and from the
relevant theory which was outlined above. The distribution of the time from the
stimulus until the channel shuts finally is simply the distribution of the sum of (a) the
first latency (mean length 1/k21 say), and (b) the length of the channel opening (mean
length 1/k12 say). Since both have been taken to be simple exponentials, as defined in
(3.12) and (3.11) respectively, this distribution is given by the convolution in (3.13).
The result, f(t), has already been given in (3.14), and illustrated in Fig. 4. It has the
form of the difference between two exponentials, which is what has been fitted to the
averages in Fig. 11.
In this particular simple case, though not in general, there is a very simple
relationship between the distribution, f(t), of the total event length, and the shape of
the averaged current. The time course of the current is given, apart from a scale
factor, by the probability that a channel is open at time t. Now a channel will be open
at time t if (a) the first latency is of length u, and (b) the channel stays open for a time
equal to or greater than t−u. The probability that a channel stays open for a time t−u
or longer, is, from (3.11), the cumulative distribution
F1(t − u) ≡ e−k12(t−u)
(7.1)
(see Chapter 6, equations 6.4-6.5), so, by an argument exactly like that used to arrive
at (3.13), the probability that a channel is open at time t is
Popen(t) =
兰 u = 0 f2(u)F1(t − u) du
u=t
(7.2)
This differs from (3.14) only by a factor of 1/k12=mean open lifetime, so
Popen(t) = f(t)/k12
(7.3)
which is, apart from its amplitude, unchanged when k12 and k21 are interchanged The
amplitudes of the two exponential components are equal and opposite, being, from
(3.14),
(7.4)
a = k21/(k21 − k12) ,
with a maximum at tmax defined in (3.15). The simulated average currents in Fig. 11
are indeed well-fitted by these values.
It is clear from this discussion that observations on the average (macroscopic)
The interpretation of single channel recordings
177
current alone cannot tell us whether the slow decay of an observed average results
from a long first latency, or from a long channel opening (see, for example, Edmonds
& Colquhoun, 1992, for experimental results). In order to distinguish between these
possibilities it is necessary to measure either the first latency distribution, or the
length of channel activations (or preferably both) independently. In practice this may
not be as easy as it sounds (a) because the channel activations are usually more
complicated than the single openings assumed here (in the case of the NMDA type of
glutamate receptor they are much more complicated), and (b) because the
measurements must be made on a patch of membrane that contains only one ion
channel molecule (or at least a known number of molecules), and this is not easy to
achieve in practice.
8. The problems of missed events and inference of mechanisms
Inferring reaction mechanisms
The discussion so far has concentrated on looking at the behaviour of specified
reaction schemes. However the most important question of all comes at an earlier
stage, namely, what is the qualitative nature of the reaction mechanism? The first
stage is usually to inspect the distributions of open times and shut times. If these
distributions can be fitted well with a sum of exponential components then the
number of components required can be taken as a (minimum) estimate of the number
of open states and the number of shut states, respectively. If the record can clearly be
divided into bursts of openings then other distributions can be investigated, for
example the distribution of the burst length, the distribution of the number of
openings per burst, and the distribution of the total open time per burst. The last two
should both have a number of components equal to the number of open states, and the
distribution of the total open time per burst is of particular interest because, unlike the
distribution of open time, it is relatively little affected by inability to detect brief
shuttings of the channel.
The next question concerns how the open and shut states are connected to each
other. The measurement of correlations between openings, bursts etc. may cast light
on this problem, as may investigation of the distributions at various times after a
voltage jump or other perturbation. In addition the presence or absence of a
component of isolated openings in the distribution of the number of openings per
burst may be informative (Colquhoun & Hawkes, 1987). However the single channel
record does not, in principle, contain sufficient information to allow all of the
connections to be worked out unambiguously. Single channel analysis, just like all
other experimental work (whether biochemical or physiological) procedes by
erecting hypotheses and then trying to demolish them with experimental results that
are incompatible with the predictions of the hypotheses.
178
D. COLQUHOUN AND A. G. HAWKES
Missed-event corrections based on a postulated mechanism
The problems that arise from the inability to detect, in practice, the briefest openings
and shuttings has already been discussed, from the practical point of view, in Chapter 6
(§10). It was pointed out there that if substantial numbers of both openings and
shuttings were missed, then it was possible to correct for this imperfection only in
cases where a kinetic mechanism could be postulated for the channel. Even then there
may be no unique solution of the problem (Colquhoun & Sigworth, 1983). Corrections
based on a postulated mechanism will be discussed next.
When openings and shuttings are missed in substantial numbers, the distributions of
the observed (inaccurate) open and shut times are no longer expected to be described
by a mixture of exponentials, as has been assumed throughout, and the theory gets a
good deal more complicated. Several approximate methods for dealing with the
problem have been suggested, e.g. Roux & Sauvé (1985), Wilson & Brown (1985),
Blatz & Magleby (1986), Ball & Sansom (1988), Milne et al. (1989) and Crouzy &
Sigworth (1990). However an exact solution to the problem (as usually formulated)
was found by Hawkes, Jalali & Colquhoun (1990) (their calculations suggest that the
best of the approximations is that of Crouzy & Sigworth). Calculation of the exact
result for short times (where it is simple), combined with use of an asymptotic
approximation (Hawkes, Jalali & Colquhoun, 1992) at longer times, allows accurate
prediction of the observed distributions for any specified mechanism and time
resolution. Such calculated distributions will be referred to as HJC distributions.
Fitting a mechanism directly to data
The ability to calculate the distribution of the quantities that are actually observed
provides a way to correct for missed events. But in addition, it also opens the way to
fitting a specified mechanism directly to the data. Previously all one could do (as in
the examples in Chapter 6) was to fit empirical mixtures of exponentials separately to
open times, shut times, burst lengths etc., which yielded values for observed time
constants and areas; the corrections for missed events, the interpretation of the results
in terms of a mechanism, and the estimation of the underlying ‘mass action’ rate
constants (see §2) for the mechanism, all had to be done retrospectively. Now that
missed events can be allowed for, it is possible to fit the HJC distributions directly to
the measured open and shut times, with the adjustable parameters in the fit not being
empirical time constants and areas, but the underlying rate constants for the model.
Since the time resolution must be specified in order to do this, it becomes particularly
important to ensure that this is known and consistent, as discussed in Chapter 6 (§5).
Simultaneous maximum likelihood fitting
In fact it may be possible to do much better than this. It is a problem with the
conventional approach to inference of mechanisms that there are many different
sources of information to be collated. For example one may have distributions of
open times, shut times, burst lengths, numbers of openings per burst, correlations, and
many others. Furthermore, the information from some sorts of fit overlaps heavily
The interpretation of single channel recordings
179
(e.g. the distribution of total open time per burst may be very similar to the
distribution of burst length) so it can be a problem to decide how many such things to
calculate.
However, once one can calculate the probability (density) for the observed open
and shut time, by means of the HJC distributions, or approximations to them, then it
becomes feasible to calculate the likelihood (see Chapter 6) of an entire single
channel record (see Horn & Lange, 1983; Horn & Vandenberg, 1984; Fredkin &
Rice, 1992). This means that, in order to do the fitting, we do not have to fit separately
all of the sorts of distribution mentioned above; instead we simply adjust the
parameters (mass action rate constants) so as to maximise the likelihood for the entire
record. This takes into account simultaneously the information from the open times,
from the shut times and from the order in which they occur (i.e. from the burst
structure and from correlations between events). We have found this to be quite
feasible on a fast PC, at least for a few thousands of observations. Further than this, it
also becomes feasible to fit several different sorts of measurement simultaneously,
for example recordings made with different concentrations of agonist (Hawkes et al.,
in preparation).
This approach can be used to judge the relative merits of alternative postulated
mechanisms. If each of the proposed mechanisms is fitted to the same data, the
relative plausibility of each mechanism can be assessed from how large its likelihood
is. This has been done, for example, by Horn & Vandenberg (1984) (without missed
event correction).
An example of fitting with missed events
An example of maximum likelihood fitting by the HJC method (to simulated single
channel data) is shown in Fig. 12.
The data in the histograms in Fig. 12 were simulated using the 5-state model for the
nicotinic acetylcholine receptor, with the rate constants and concentration used by
Colquhoun & Hawkes (1982) as shown in (8.1). The numbers are transition rates in
s−1 (see discussion following 2.1).
(8.1)
A resolution of 100 µs (for both open and shut times) was then imposed on the
180
D. COLQUHOUN AND A. G. HAWKES
simulated sequence of open and shut times (see Chapter 6), to produce a sequence of
apparent open and shut times from which the histograms were constructed. These
histograms were not themselves fitted, but the likelihood of the entire sequence of
apparent open and shut times was calculated, as outlined above, and the free
A
256
true pdf (scaled)
Frequency (square root scale)
196
144
100
64
36
16
4
0
0.01
0.1
484
1
10
Apparent open time (ms) (log scale)
100
true pdf (scaled)
B
Frequency (square root scale)
400
324
256
196
144
100
64
36
16
4
0
0.01
0.1
1
10
100
1000
Apparent shut time (ms) (log scale)
10000
100000
Fig. 12. Example of the effect of missed events. The 5-state mechanism in (8.1) was used to
simulate 10235 channel transitions. A resolution of 100 µs was imposed on the simulated
channels resulting in 3419 resolved intervals, from which the histograms of apparent open
tines (A), and apparent shut times (B) were constructed. The continuous lines are the HJC
distributions calculated for a resolution of 100 µs with rate constants as in (8.1); the exact
solution is used for intervals up to 300 µs, and the asymptotic solution thereafter. The dashed
lines show the true open and shut time distributions.
The interpretation of single channel recordings
181
parameters adjusted (by the simplex method) to maximize this likelihood. The
estimates of the transition rates, from this single fitting procedure, were close to the
values shown in (8.1), as hoped. The true values from (8.1) were then used to
calculate the HJC distributions for the apparent open times and shut times. These are
shown as continuous lines superimposed on the histograms in Fig. 12. It can be seen
that they fit the ‘observations’ quite well. The same values were then used to calculate
the predicted open and shut time distributions by the standard methods (e.g.
Colquhoun & Hawkes, 1982), with no allowance for missed events. These
distributions, which are what would be expected if we had perfect time resolution, are
shown as dashed lines in Fig. 12A,B.
The open time distribution in Fig. 12A shows that the openings appear to be
considerably longer than their true values, because many brief shuttings are below the
100 µs resolution, and therefore not detected. The true fast shut time component has
τ=52.6 µs, so 85% of this component is missed. With perfect resolution the open time
distribution would, with the rate constants in (8.1), have two exponential components
with τ=2.0 ms (93% of area) and τ=0.33 ms (7% of area). The asymptotic HJC
distribution of apparent open times, on the other hand, has components with τ=6.1 ms
(81% of area) and 0.33 ms (19% of area).
The shut time distribution in Fig. 12B shows, in contrast, that the distribution of
apparent shut times is close to the true distribution, except of course that there are no
shut times below 100 µs: this happens because, in this particular example, relatively
few openings are below 100 µs. This shows, incidentally, that for results of this sort,
the number of missed shuttings can be estimated quite accurately by extrapolating the
shut time distribution to zero length. Thus the crude form of retrospective missed
event correction used by Colquhoun & Sakmann (1985) (see also Chapter 6) should
have been reasonably accurate.
Problems: the number of channels
Perhaps the biggest single problem that hinders the interpretation of single channel
records stems from the fact that one rarely knows how many functional channels were
present in the membrane patch from which the record was made. Shut time
distributions, including first latency distributions, must obviously depend on the
number of channels that are present.
If it is observed, at any time during the recording, that more than one channel is
open, then the patch must contain more than one channel, but it is notoriously
difficult to estimate how many there are, at least when Popen is low (see Horn, 1991,
and Chapter 6, §5). A method based on the length of runs of single openings has been
proposed (Colquhoun & Hawkes, 1990), which may be useful if the burst structure of
the data is not too complex.
Perhaps the most common way of managing this problem is to confine attention
only to subsections of the record that can, with confidence, be attributed to the activity
of only one channel. Such subsections may consist of individual bursts (e.g. at low
agonist concentrations), or may consist of much longer clusters of openings when
182
D. COLQUHOUN AND A. G. HAWKES
Popen is high (see Chapter 6). This means, of course, that information from longer shut
periods is lost. This approach can be incorporated into maximum likelihood fitting
with HJC distributions (Hawkes, Jalali, & Colquhoun, in preparation).
Problems: indeterminacy of parameters
It is unlikely that any individual experimental record will contain enough
information to allow estimation of all the rate constants in the postulated
mechanism. This problem can be minimized if several sorts of experiment (e.g.
equilibrium records with different ligand concentrations, jump experiments, Popen
curves etc.) can be analysed simultaneously, as described above. But in practice it
will often be necessary to estimate some rate constants form one sort of experiment,
and then to treat these values as fixed while estimating others from a different sort of
experiment (see, for example, Colquhoun & Sakmann, 1985; Sine, Claudio &
Sigworth, 1990).
Problems: indeterminacy of mechanisms
It is well-known that kinetic mechanisms are not unique. Whatever mechanism is
proposed, it will virtually always be possible to find another mechanism that fits the
experimental results just as well. Some classes of mechanism predict results that are
too similar for them to be distinguished in practice. Worse still, some classes of
mechanism are indistinguishable in principle; this has been discussed very elegantly
by Kienker (1989). Nevertheless, combination of different sorts of biophysical
experiment (as in the last paragraph), together with biochemical and structural
information, can do much to reduce the ambiguity.
These problems lead some people to be pessimistic about the possibility of
interpreting experiments in terms of physical mechanisms. One extreme reaction is to
point out that proteins have an essentially infinite number of conformations (which is,
no doubt, true), and must therefore be analysed by fractal methods, with consequent
denial that anything can be gleaned about the physical mechanisms involved (see. for
example, Korn & Horn, 1988). We think that it is perverse to suggest that
experimental results have not been able to cast light on topics such as mechanisms of
ion channel block, or the fine structure of bursts; the proposed mechanisms are
doubtless approximations, but they are not baseless.
Appendix 1. Derivation of the shut time distribution for the
Castillo-Katz mechanism
The Castillo-Katz mechanism defined in (4.1) will be written again here, for easy
reference.
The interpretation of single channel recordings
vacant
occupied
k+1
β
k−1
α
A + R [ AR[ AR*
state number:
183
3
2
shut
(A1.1)
1
open
The derivation is fairly lengthy, even for this mechanism with only three states.
This complexity results, to a great extent, from the fact that the two shut states can
interconvert directly (unlike the simple channel block mechanism), so that there may
be any number of oscillations between them during any individual shut period. It can
easily be imagined that derivations such as that to follow are not generally feasible for
mechanisms with more than three states, and even with three states the complexity is
great enough to provide a considerable incentive to program the quite general matrix
results (Colquhoun & Hawkes, 1982) which will give numerical results for any
mechanism.
First, notice that when the channel is at equilibrium, every shut period follows an
opening, and so must start with a sojourn in the intermediate complex, AR (state 2).
After leaving AR the channel may reopen immediately, or it may go through any
number of AR[R oscillations before reopening. Furthermore every shut period must
end in state 2 (AR) also, because the shut period ends when an opening occurs, and
opening is possible only from state 2.
We start by defining the probability, denoted P22(t), that the channel starts in state
2, stays shut (state 2 or 3) throughout the time period from 0 to t, and is in state 2 at t.
The shut lifetime will be t if, having stayed shut throughout the time from 0 to t, the
channel then opens. The probability that a channel opens during a small time
interval, ∆t, is, by analogy with (2.10), β∆t. The probability density function for all
shut times, fs(t) say, can be found by multiplying these two probabilities (the
probabilities are independent, because the probability of a transition from state 2 to
state 1 in the interval between t and t+∆t is independent of the previous history of the
process): it is, therefore, the limiting value, for very small ∆t, of the following
expression.
fs(t) = Prob[channel shut from 0 to t and opens during t, t+∆t]/∆t
= P22(t)β
(A1.2)
We can now get an expression for P22(t), in the following way. The channel will be
in state 2 (AR) at time t+∆t if either (1) it was in state 2 at time t, and fails to leave
state 2 during ∆t, or (2) it was in state 3 (R) at time t, and moves from 2 to 3 (i.e. binds
a ligand molecule) during ∆t. To evaluate the first of these contingencies, we note that
the probability of leaving state 2 (for either state 3 or state 1) during ∆t is (β+k−1)∆t,
so the probability of not leaving is 1 minus this quantity. To evaluate the second term
184
D. COLQUHOUN AND A. G. HAWKES
we must first define P23(t) as the probability that a channel that is in state 2 at t=0 then
remains shut throughout the time from 0 to t and is in state 3 at t; the probability of a
transition from 3 to 2 in ∆t is then k+1xA∆t. We can now assemble all the possibilities
to write
P22(t + ∆t) = P22(t)[1 − (β + k−1)∆t] + P23(t)k+1xA∆t .
(A1.3)
P22(t+∆t) − P22(t)
–––––––––––––––– = − P22(t)(β + k−1) + P23(t)k+1xA
∆t
(A1.4)
Thus
and so, letting ∆t→0, we obtain
dP22(t)
–––––––– = − P22(t)(β + k−1) + P23(t)k+1xA
dt
(A1.5)
This equation contains two unknowns, P22(t) and P23(t). To solve this problem we
must go through an exactly analogous argument to obtain a second equation, viz.
dP23(t)
–––––––– = P22(t) k−1 − P23(t)k+1xA .
dt
(A1.6)
These two equations allow the two unknowns to be evaluated. To achieve this we
eliminate P23(t) from the equations by using (A1.5) to obtain an expression for P23(t),
which is then substituted into (A1.6). The result is a second order equation that
involves only P22(t), thus
dP22(t)
d2P22(t)
–––––––– + (β + k−1 + k+1xA) ––––––– + k+1xAβP22(t) = 0
dt
dt2
(A1.7)
If we define the (constant) coefficients, b and c, as
b = λ1 + λ2 = β + k−1 + k+1xA
(A1.8)
c = λ1λ2 = k+1xAβ
(A1.9)
then (A1.7) can be written in the form
dP22(t)
d2P22(t)
––––––– + b ––––––– + cP22(t) = 0
2
dt
dt
(A1.10)
We are now in a position to obtain expressions for the observed time constants. A
solution of (A1.10) is needed, and we proceed, in the irritating manner of
mathematicians, to propose, for no apparent reason, that there is a particular solution
of (A1.10) that has the form P22(t)=e−λt. If this is the case then dP22(t)/dt=−λe−λt, and
d2P22(t)/dt2=λ2e−λt. Substitution of these into (A1.10) gives:
The interpretation of single channel recordings
λ2e−λt − bλe−λt + ce−λt = 0
185
(A1.11)
This result must be true at all times, including at t=0. Putting t=0 in (A1.11) gives
λ2 − bλ + c = 0
(A1.12)
This is a quadratic equation and therefore there will generally be two values of the
observed rate constants, λ1 and λ2 say, that satisfy it, and these can be found as the
two solutions of (A1.12), namely
λ1, λ2 = 0.5共b ± √b2 − 4c兲
(A1.13)
2c
λ1, λ2 = –––––––––––– ,
b⫿√b2 − 4c
(A1.14)
or
where the coefficients, b and c, were defined in (A1.8) and (A1.9). Notice that these
coefficients are rather simpler than those given in (5.4) and (5.5) for the macroscopic
relaxation, because we are now dealing only with the shut states. They no longer
involve all of the fundamental rate constants (α does not appear).
To complete the distribution of shut times, we must next find the areas, a1 and a2,
of the two components. The p.d.f. has the form (see Chapter 6)
fs(t) = βP22(t) = a1λ1e−λ1t + a2λ2e−λ2t
(A1.15)
where a2=1−a1, because the total are must be 1. Thus, from (A1.2) and (A1.15),
dfs(t)
––––––
dt
t=0
βdP22(t)
= ––––––
dt
= − a1λ12 − a2λ22
t=0
= a1(λ22 − λ12 ) − λ22
= −β(β + k−1)
(A1.16)
The last line follows from (A1.5) because P22(0)=1 (state 2 cannot be left in zero
time), and P23(0)=0 (cannot get from state 2 to 3 in zero time). From this, together
with the fact that λ1+λ2=β+k−1+k+1xA and λ1λ2=k+1xAβ (see A1.8 and A1.9), we find,
after some manipulation, that the areas of the two components are given by
β(k+1xA − λ1)
a1 = –––––––––––– , and a2 = 1 − a1 .
(λ2 − λ1)λ1
(A1.17)
This completes the derivation of the distribution of all shut times. The same result
was derived, in a quite different way, by Colquhoun & Hawkes (1981), but if all that
is needed is numerical values for the time constants and areas of the components it is
obviously much easier to use the entirely general result (Colquhoun & Hawkes,
186
D. COLQUHOUN AND A. G. HAWKES
1982), rather than go through a derivation like that above for every mechanism of
interest.
Approximations for the time constants. If b2 » 4c then it follows from (A1.13) that
one of the observed rate constants will be much larger than the other, say λ2 » λ1. This
will be the case, for example, when the ligand concentration, xA, is very low, so from
(A1.8), we find that the faster time constant is approximately
1
τ2 = 1/λ2 ≈ –––––––
β + k−1
(A1.18)
The right hand side of this is simply the mean lifetime of a single sojourn in the
intermediate complex AR (state 2). This is, therefore, an example of a case where,
contrary to the general rule, a physical interpretation can be placed on an observed
time constant, as an approximation. In this case the sojourns in AR that occur as the
channel oscillates between AR[AR*, can be seen (approximately) as the fast
component of the distribution of all shut times.
References
ADAMS, P. R. (1976). Drug blockade of open end-plate channels. J. Physiol. 260, 531-552.
ADAMS, P. R. & SAKMANN, B. (1978). Decamethonium both opens and blocks endplate channels.
Proc. Natl. Acad. Sci. U.S.A. 75, 2994-2998.
ALDRICH, R. W., COREY, D. P. & STEVENS, C. F. (1983). A reinterpretation of mammalian sodium
channel gating based on single channel recording. Nature 306, 436-441.
ANDERSON, C. R. & STEVENS, C. F. (1973). Voltage clamp analysis of acetylcholine produced endplate current fluctuations at frog neuromuscular junction. J. Physiol. 235, 655-691.
BALL, F. G. & SANSOM, M. S. P. (1988). Aggregated Markov processes incorporating time interval
omission. Adv. Appl. Prob. 20, 546-572.
BALL, F. G. & SANSOM, M. S. P. (1988). Single channel autocorrelation functions. The effects of time
interval omission. Biophys. J. 53, 819-832.
BLATZ, A. L. & MAGLEBY, K. L. (1986). Correcting single channel data for missed events. Biophys.
J. 49, 967-980.
BLATZ, A. L. & MAGLEBY, K. L. (1989). Adjacent interval analysis distinguishes among gating
mechanisms for the fast chloride channel from rat skeletal muscle. J. Physiol. 410, 561-585.
CASTILLO, J. del & KATZ, B. (1957). Interaction at end-plate receptors between different choline
derivatives. Proc. Roy. Soc. B 146, 369-381.
COLQUHOUN, D. (1971). Lectures on Biostatistics. Oxford: Clarendon Press.
COLQUHOUN, D., DREYER, F. & SHERIDAN, R. E. (1979). The actions of tubocurarine at the frog
neuromuscular junction. J. Physiol. 293, 247-284.
COLQUHOUN, D. & HAWKES, A. G. (1977). Relaxation and fluctuations of membrane currents that
flow through drug-operated ion channels. Proc. Roy. Soc. B 199, 231-262.
COLQUHOUN, D. & HAWKES, A. G. (1981). On the stochastic properties of single ion channels.
Proc. Roy. Soc. B 211, 205-235.
COLQUHOUN, D. & HAWKES, A. G. (1982). On the stochastic properties of bursts of single ion
channel openings and of clusters of bursts. Phil. Trans. Roy. Soc. B 300, 1-59.
COLQUHOUN, D. & HAWKES, A. G. (1983). The principles of the stochastic interpretation of ion
channel mechanisms. In Single Channel Recording (ed. B. Sakmann & E. Neher). New York: Plenum
Press.
COLQUHOUN, D. & HAWKES A. G. (1987). A note on correlations in single channel records. Proc.
Roy. Soc. B 230, 15-52.
COLQUHOUN, D. & HAWKES, A. G. (1990). Stochastic properties of ion channel openings and bursts
The interpretation of single channel recordings
187
in a membrane patch that contains two channels: evidence concerning the number of channels present
when a record containing only single openings is observed. Proc. Roy. Soc. London B 240, 453-477.
COLQUHOUN, D., JONAS, P. & SAKMANN, B. (1992). Action of brief pulses of glutamate on
AMPA/kainate receptors in patches from different neurones of rat hippocampal slices. J. Physiol. 458,
261-287.
COLQUHOUN, D. & OGDEN, D. C. (1986). States of the acetylcholine receptor: enumeration
characteristics and structure. In Nicotinic Acetylcholine Receptor: structure and function (ed. A.
Maelicke). Berlin: Springer-Verlag. pp. 197-232.
COLQUHOUN, D., OGDEN. D. C. & CACHELIN, A. B. (1986). Mode of action of agonists on
nicotinic receptors. In Ion Channels in Neural Membranes (ed. J. M. Ritchie, R. D. Keynes &
L.Bolis), pp. 255-273. New York: A. R. Liss.
COLQUHOUN, D. & SAKMANN, B. (1981). Fluctuations in the microsecond time range of the current
through single acetylcholine receptor ion channels. Nature 294, 464-466.
COLQUHOUN, D. & SAKMANN, B. (1983). Bursts of openings in transmitter-activated ion channels.
In Single Channel Recording (ed. B. Sakmann & E. Neher). New York: Plenum Press.
COLQUHOUN, D. & SAKMANN, B. (1985). Fast events in single-channel currents activated by
acetylcholine and its analogues at the frog muscle end-plate. J. Physiol. 369, 501-557.
COLQUHOUN, D . & SHERIDAN R. E. (1981) . The modes of action of gallamine. Proc. Roy. Soc.
London B 211, 181-203.
CROUZY, S. C. & SIGWORTH, F. J. (1990). Yet another approach to the dwell-time omission problem
of single-channel analysis. Biophys. J. 58, 731-743.
EDMONDS, B. & COLQUHOUN, D. (1992). Rapid decay of averaged single-channel NMDA receptor
activations recorded at low agonis concentration. Proc. Roy. Soc. London B, 250, 279-286.
FRANKE, CH., HATT, H., PARNAS, H. & DUDEL, J. (1991). Kinetic constants of the acetylcholine
(ACh) receptor reaction deduced from the rise in open probability after steps in ACh concentration.
Biophys. J. 60, 1008-1016.
FREDKIN, D. R., MONTAL, M. & RICE, J. A. (1985). Identification of aggregated Markovian models:
application to the nicotinic acetylcholine receptor. In: Proceedings of the Berkeley Conference in
honor of Jerzy Neyman and Jack Kiefer, vol. I (ed. L. M. Le Carn & R. A. Olshen). pp. 269-289.
Wadsworth Press.
FREDKIN, D. R. & RICE, J. A. (1992). Maximum likelihood estimation and identification directly from
single-channel recordings. Proc. Roy. Soc. London B 249, 125-132.
GIBB, A. J. & COLQUHOUN, D. (1992). Activation of NMDA receptors by L-glutamate in cells
dissociated from adult rat hippocampus. J. Physiol. 456, 143-179.
HAMILL, O. P., MARTY, A., NEHER, E., SAKMANN, B. & SIGWORTH, F. J. (1981). Improved
patch-clamp techniques for high-resolution current recording from cells and cell-free membrane
patches. Pflugers Archiv. 391, 85-100.
HAWKES, A. G., JALALI, A. & COLQUHOUN, D. (1990). The distributions of the apparent open
times and shut times in a single channel record when brief events can not be detected. Phil. Trans.
Roy. Soc. London A 332, 511-538.
HAWKES, A.G., JALALI, A. & COLQUHOUN, D. (1992). Asymptotic distributions of apparent open
times and shut times in a single channel record allowing for the omission of brief events. Phil. Trans.
Roy. Soc. London B 337, 383-404.
HORN, R. (1991). Estimating the number of channels in patch recordings. Biophys. J. 60, 433-439.
HORN, R. (1984). Gating of channels in nerve and musclc: a stochastic approach. In Ion Channels:
Molecular and Physiological Aspects (ed. W. D. Stein). pp. 53-97. New York: Academic Press.
HORN, R. & LANGE, K. (1983). Estimating kinetic constants from single channel data. Biophys. J. 43,
207-223.
HORN, R. & VANDENBURG, C. A. (1984). Statistical properties of single sodium channels. J. Gen.
Physiol. 84, 505-534.
JACKSON, M. B., WONG, B. S., MORRIS, C. E., LECAR, H. & CHRISTIAN, C. N. (1983).
Successive openings of the same acetylcholine receptor channels are correlated in open time. Biophys.
J. 42, 109-114.
KIENKER, P. (1989). Equivalence of aggregated Markov models of ion-channel gating. Proc. Roy. Soc.
London B 236, 269-309.
KORN, J. S. & HORN, R. (1988). Statistical discrimination of fractal and Markov models of singlechannel gating. Biophys. J. 54, 871-877.
188
D. COLQUHOUN AND A. G. HAWKES
LABARCA, P., RICE, J. A., FREDKIN, D. R. & MONTAL, M. (1985). Kinetic analysis of channel
gating: application to the cholinergic receptor channel and the chloride channel from Torpedo
californica. Biophys. J. 47, 469-478.
LIU, Y. & DILGER, J. P. (1991). Opening rate of acetylcholine receptor channels. Biophys. J. 60, 424432.
MAGLEBY, K. L. & WEISS, D. S. (1990). Identifying kinetic gating mechanisms for ion channels by
using two-dimensional distributions of simulated dwell times. Proc. Roy. Soc. London B 241, 220228.
McMANUS, O. B., BLATZ, A. L. & MAGLEBY, K. L. (1985). Inverse relationship of the durations of
open and shut intervals for Cl and K channels. Nature 317, 625-627.
MILNE, R. K., YEO, G. F., EDESON, R. O. & MADSEN, B. W. ( 1989). Estimation of single channel
kinetic parameters from data subject to limited time resolution. Biophys. J. 55, 673-676.
NEHER, E. (1983). The charge carried by single-channel currents of rat cultured muscle cells in the
presence of local anaesthetics. J. Physiol. 339, 663-678.
NEHER, E. & STEINBACH, J. H. (1978). Local anaesthetics transiently block currents through single
acetylcholine-receptor channels. J. Physiol. 277, 153-176.
OGDEN, D. C. & COLQUHOUN, D. (1985). Ion channel block by acetylcholine, carbachol and
suberyldicholine at the frog neuromuscular junction. Proc. Roy. Soc. London B 225, 329- 355.
OGDEN, D. C., SIEGELBAUM, S. A. & COLQUHOUN, D. (1981). Block of acetylcholine- activated
ion channels by an uncharged local anaesthetic. Nature 289, 596-599.
ROUX, B. & SAUVE, R. (1985). A general solution to the time interval omission problem applied to
single channel analysis. Biophys. J. 48, 149-158.
SAKMANN, B., PATLAK, J. & NEHER, E. (1980). Single acetylcholine-activated channels show
burst-kinetics in presence of desensitizing concentrations of agonist. Nature 286, 72-73.
SAKMANN, B. & TRUBE, G. (1984). Voltage-dependent inactivation of inward-rectifying single
channel currents in the guinea-pig heart cell membrane. J. Physiol. 347, 659-683.
SINE, S. M. & STEINBACH, J. H. (1986). Activation of acetylcholine receptors on clonal BC3H-1 cells
by low concentrations of agonist. J. Physiol. 373, 129-162.
SINE, S. M., CLAUDIO, T. & SIGWORTH, F. J. (1990). Activation of Torpedo acetylcholine receptors
expressed in mouse fibroblasts: single channel current kinetics reveal distinct agonist binding
affinities. J. Gen. Physiol. 96, 395-437.
THOMPSON, S.P. (1965). Calculus Made Easy. London: Macmillan.
WILSON, D. L. & BROWN, A. M. (1985). Effect of limited interval resolution on single channel
measurements with application to Ca channels. IEEE Trans. Biomed. Eng. 32, 780-797.